95         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()  local
102         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
110         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
118         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
126         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
134         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
142         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
150         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
158         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
166         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
181         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
189         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
197         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
205         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
213         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
221         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
229         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
237         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
245         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
253         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
261         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
269         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
284         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c1x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
292         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
300         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
308         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
316         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
324         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
336         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
341         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
346           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
356             vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
357             vo1c0123 = vo1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
362           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
365             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
382         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()  local
389         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
397         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
405         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
413         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
421         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
429         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
437         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
445         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
453         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
468         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
476         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
484         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
492         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
500         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
508         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
516         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
524         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
532         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
538         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
543         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
548           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
558             vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
559             vo1c0123 = vo1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
564           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
567             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()