94         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()  local
99         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
104         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
109         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
114         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
119         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
124         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
129         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
134         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
139         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
151         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
156         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
161         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
166         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
171         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
176         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
181         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
186         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
191         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
196         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
201         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
206         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
218         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c1x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
223         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
228         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
233         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
238         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
243         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
253         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
256         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
259           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
267           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
270             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
286         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()  local
291         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
296         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
301         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
306         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
311         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
316         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
321         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
326         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
331         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
343         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
348         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
353         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
358         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
363         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
368         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
373         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
378         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
383         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
387         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
390         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
393           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
401           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
404             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()