246         const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
249         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
250         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
254         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_high_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
255         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
650         const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
653         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
654         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
936         const float32x4_t vk11c0x0123 = vld1q_f32(w + 88);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
939         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
940         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()