105         float32x4_t vo1x0 = vo0x0;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
119         vo1x0 = vmlaq_lane_f32(vo1x0, vk00c0, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
126         vo1x0 = vmlaq_lane_f32(vo1x0, vk10c0, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
133         vo1x0 = vmlaq_lane_f32(vo1x0, vk20c0, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
147         vo1x0 = vmlaq_lane_f32(vo1x0, vk00c1, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
154         vo1x0 = vmlaq_lane_f32(vo1x0, vk10c1, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
161         vo1x0 = vmlaq_lane_f32(vo1x0, vk20c1, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
168         vo1x0 = vmlaq_lane_f32(vo1x0, vk00c2, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
175         vo1x0 = vmlaq_lane_f32(vo1x0, vk10c2, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
182         vo1x0 = vmlaq_lane_f32(vo1x0, vk20c2, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
189         vo1x0 = vmlaq_lane_f32(vo1x0, vk01c0, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
196         vo1x0 = vmlaq_lane_f32(vo1x0, vk11c0, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
203         vo1x0 = vmlaq_lane_f32(vo1x0, vk21c0, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
210         vo1x0 = vmlaq_lane_f32(vo1x0, vk01c1, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
217         vo1x0 = vmlaq_lane_f32(vo1x0, vk11c1, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
224         vo1x0 = vmlaq_lane_f32(vo1x0, vk21c1, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
238         vo1x0 = vmlaq_lane_f32(vo1x0, vk01c2, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
245         vo1x0 = vmlaq_lane_f32(vo1x0, vk11c2, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
252         vo1x0 = vmlaq_lane_f32(vo1x0, vk21c2, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
259         vo1x0 = vmlaq_lane_f32(vo1x0, vk02c0, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
266         vo1x0 = vmlaq_lane_f32(vo1x0, vk12c0, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
273         vo1x0 = vmlaq_lane_f32(vo1x0, vk22c0, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
280         vo1x0 = vmlaq_lane_f32(vo1x0, vk02c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
287         vo1x0 = vmlaq_lane_f32(vo1x0, vk12c1, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
294         vo1x0 = vmlaq_lane_f32(vo1x0, vk22c1, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
301         vo1x0 = vmlaq_lane_f32(vo1x0, vk02c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
308         vo1x0 = vmlaq_lane_f32(vo1x0, vk12c2, vget_low_f32(vi3x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
315         vo1x0 = vmlaq_lane_f32(vo1x0, vk22c2, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
326         vo1x0 = vmaxq_f32(vo1x0, vmin);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
331         vo1x0 = vminq_f32(vo1x0, vmax);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
336         const float32x4x2_t vo1c0123 = vzipq_f32(vo1x0, vo1x1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
352         float32x4_t vo1x0 = vo0x0;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
366         vo1x0 = vmlaq_lane_f32(vo1x0, vk00c0, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
375         vo1x0 = vmlaq_lane_f32(vo1x0, vk10c0, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
384         vo1x0 = vmlaq_lane_f32(vo1x0, vk20c0, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
407         vo1x0 = vmlaq_lane_f32(vo1x0, vk00c1, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
414         vo1x0 = vmlaq_lane_f32(vo1x0, vk10c1, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
421         vo1x0 = vmlaq_lane_f32(vo1x0, vk20c1, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
428         vo1x0 = vmlaq_lane_f32(vo1x0, vk00c2, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
435         vo1x0 = vmlaq_lane_f32(vo1x0, vk10c2, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
442         vo1x0 = vmlaq_lane_f32(vo1x0, vk20c2, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
449         vo1x0 = vmlaq_lane_f32(vo1x0, vk01c0, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
458         vo1x0 = vmlaq_lane_f32(vo1x0, vk11c0, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
467         vo1x0 = vmlaq_lane_f32(vo1x0, vk21c0, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
476         vo1x0 = vmlaq_lane_f32(vo1x0, vk01c1, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
485         vo1x0 = vmlaq_lane_f32(vo1x0, vk11c1, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
494         vo1x0 = vmlaq_lane_f32(vo1x0, vk21c1, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
517         vo1x0 = vmlaq_lane_f32(vo1x0, vk01c2, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
524         vo1x0 = vmlaq_lane_f32(vo1x0, vk11c2, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
531         vo1x0 = vmlaq_lane_f32(vo1x0, vk21c2, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
539           vo1x0 = vmlaq_lane_f32(vo1x0, vk02c0, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
544           vo1x0 = vmlaq_lane_f32(vo1x0, vk12c0, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
549           vo1x0 = vmlaq_lane_f32(vo1x0, vk22c0, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
554           vo1x0 = vmlaq_lane_f32(vo1x0, vk02c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
559           vo1x0 = vmlaq_lane_f32(vo1x0, vk12c1, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
564           vo1x0 = vmlaq_lane_f32(vo1x0, vk22c1, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
569           vo1x0 = vmlaq_lane_f32(vo1x0, vk02c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
574           vo1x0 = vmlaq_lane_f32(vo1x0, vk12c2, vget_low_f32(vi3x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
579           vo1x0 = vmlaq_lane_f32(vo1x0, vk22c2, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
583         vo1x0 = vmaxq_f32(vo1x0, vmin);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
588         vo1x0 = vminq_f32(vo1x0, vmax);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
595           const float32x4x2_t vo1c0123 = vzipq_f32(vo1x0, vo1x1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
609           vst1q_lane_f32(o1c0, vo1x0, 0); o1c0 += 1;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
610           vst1q_lane_f32(o1c1, vo1x0, 1); o1c1 += 1;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
611           vst1q_lane_f32(o1c2, vo1x0, 2); o1c2 += 1;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
612           vst1q_lane_f32(o1c3, vo1x0, 3); o1c3 += 1;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()