94         float32x4_t vo0c4567 = vld1q_f32(w + 4);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()  local
96         float32x4_t vo1c4567 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
103         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
111         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
119         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
127         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
135         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
143         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
151         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
159         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
167         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
175         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
183         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
191         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
206         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
214         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c1x4567, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
222         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
230         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c2x4567, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
238         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
246         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
254         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
262         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
270         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
278         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
286         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
294         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
309         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c2x4567, vi0x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
317         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
325         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
337         vo0c4567 = vmaxq_f32(vo0c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
342         vo0c4567 = vminq_f32(vo0c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
350           vst1q_f32(o0 + 4, vo0c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
359             vo0c0123 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
381         float32x4_t vo0c4567 = vld1q_f32(w + 4);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()  local
383         float32x4_t vo1c4567 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
390         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
398         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
406         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
414         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
422         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
430         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
438         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
446         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
454         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
462         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
470         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
478         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
493         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
501         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c1x4567, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
509         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
517         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c2x4567, vi0x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
525         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
533         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
539         vo0c4567 = vmaxq_f32(vo0c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
544         vo0c4567 = vminq_f32(vo0c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
552           vst1q_f32(o0 + 4, vo0c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
561             vo0c0123 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()