91         float32x4_t vo0x0c0123 = vld1q_f32(w);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()  local
93         float32x4_t vo1x0c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
95         float32x4_t vo0x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
97         float32x4_t vo1x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
110         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c0x0123, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
123         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
136         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
149         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
162         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
175         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
195         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c2x0123, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
208         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
221         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
234         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c0x0123, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
247         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
260         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
273         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c1x0123, vget_low_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
286         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c1x0123, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
299         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
312         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk01c2x0123, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
325         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk11c2x0123, vget_low_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
338         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
358         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c0x0123, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
371         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c0x0123, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
384         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
397         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c1x0123, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
410         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
423         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
436         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
449         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
462         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c2x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
481         vo0x0c0123 = vmaxq_f32(vo0x0c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
491         vo0x0c0123 = vminq_f32(vo0x0c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
505           vst1q_f32(o0, vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
526             vst1q_f32(o0_tmp, vo0x0c0123); o0_tmp += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
527             vo0x0c0123 = vo0x0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
529           float32x2_t vo0x0c01 = vget_low_f32(vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
542             vo0x0c01 = vget_high_f32(vo0x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()