92         float32x4_t vo0c4567 = vld1q_f32(w + 4);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()  local
94         float32x4_t vo1c4567 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
101         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
109         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
117         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
125         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
133         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
141         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
149         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
157         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
165         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
173         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
181         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
189         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
204         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
212         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c1x4567, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
220         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
228         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c2x4567, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
236         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
244         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
252         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
260         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
268         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
276         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
284         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
292         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
307         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c2x4567, vi0x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
315         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
323         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
337         vo0c4567 = vmaxq_f32(vo0c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
342         vo0c4567 = vminq_f32(vo0c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
350           vst1q_f32(o0 + 4, vo0c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
359             vo0c0123 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
381         float32x4_t vo0c4567 = vld1q_f32(w + 4);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()  local
383         float32x4_t vo1c4567 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
390         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
398         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
406         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
414         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
422         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
430         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
438         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
446         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
454         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
462         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
470         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
478         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
493         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
501         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c1x4567, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
509         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
517         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk01c2x4567, vi0x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
525         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
533         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
541         vo0c4567 = vmaxq_f32(vo0c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
546         vo0c4567 = vminq_f32(vo0c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
554           vst1q_f32(o0 + 4, vo0c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
563             vo0c0123 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()