96         float32x4_t vo1x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
111         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
119         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
127         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
142         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
150         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
158         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
166         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
174         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
182         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
190         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
198         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
206         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c0x0123, vget_high_f32(vi4x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
214         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c1x0123, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
222         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
230         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
245         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
253         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
261         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c2x0123, vget_low_f32(vi4x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
269         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c0x0123, vget_low_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
277         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c0x0123, vget_low_f32(vi3x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
285         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
293         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c1x0123, vget_high_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
301         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
309         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
317         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
325         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
333         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c2x0123, vget_high_f32(vi4x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
346         vo1x1c0123 = vmaxq_f32(vo1x1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
352         vo1x1c0123 = vminq_f32(vo1x1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
360           vst1q_f32(o1, vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
370           float32x2_t vo1x1c01 = vget_low_f32(vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
373             vo1x1c01 = vget_high_f32(vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()