232         const float32x4_t vk01c2x0123 = vld1q_f32(w + 64);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
234         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
235         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
237         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
238         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
481         const float32x4_t vk01c2x0123 = vld1q_f32(w + 64);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
483         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
484         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
672         const float32x4_t vk01c2x0123 = vld1q_f32(w + 64);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
674         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk01c2x0123, vi0x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
675         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()