93         float32x4_t vo0x0c0123 = vld1q_f32(w);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
94         float32x4_t vo1x0c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
95         float32x4_t vo0x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
96         float32x4_t vo1x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
107         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
115         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
123         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
131         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
139         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
147         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
162         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c2x0123, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
170         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
178         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
186         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
194         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
202         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
210         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c1x0123, vget_low_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
218         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
226         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
234         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
242         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
250         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
265         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c0x0123, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
273         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c0x0123, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
281         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
289         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c1x0123, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
297         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
305         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
313         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
321         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
329         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c2x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
342         vo0x0c0123 = vmaxq_f32(vo0x0c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
348         vo0x0c0123 = vminq_f32(vo0x0c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
357           vst1q_f32(o0, vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
367           float32x2_t vo0x0c01 = vget_low_f32(vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
380             vo0x0c01 = vget_high_f32(vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()