95         float32x4_t vo0x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
110         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
118         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
126         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
134         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c1x0123, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
142         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
150         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
165         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c2x0123, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
173         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c2x0123, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
181         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c2x0123, vget_low_f32(vi2x2), 0);  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()
197         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
205         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c0x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
213         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c1x0123, vget_high_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
221         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
229         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0);  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()
245         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 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()
268         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c0x0123, vget_low_f32(vi0x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
276         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c0x0123, vget_low_f32(vi1x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
284         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
292         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c1x0123, vget_low_f32(vi0x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
300         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
308         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
316         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
324         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
332         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c2x0123, vget_high_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
345         vo0x1c0123 = vmaxq_f32(vo0x1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
351         vo0x1c0123 = vminq_f32(vo0x1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
362           vst1q_f32(o0, vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
369           float32x2_t vo0x1c01 = vget_low_f32(vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
375             vo0x1c01 = vget_high_f32(vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()