91         float32x4_t vo0x0c0123 = vld1q_f32(w);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()  local
92         float32x4_t vo1x0c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
93         float32x4_t vo0x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
94         float32x4_t vo1x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
105         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
113         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
121         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
136         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
144         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
152         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
160         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c2x0123, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
168         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c2x0123, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
176         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
184         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_low_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
192         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
200         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
208         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c1x0123, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
216         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
224         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
239         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
247         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
255         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
263         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c0x0123, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
271         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c0x0123, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
279         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
287         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c1x0123, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
295         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
303         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
311         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
319         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
327         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c2x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
342         vo0x0c0123 = vmaxq_f32(vo0x0c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
348         vo0x0c0123 = vminq_f32(vo0x0c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
357           vst1q_f32(o0, vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
367           float32x2_t vo0x0c01 = vget_low_f32(vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
380             vo0x0c01 = vget_high_f32(vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()