95         float32x4_t vo1x0c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
113         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
126         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
139         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
152         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
165         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
178         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
198         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
211         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
224         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
237         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
250         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
263         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
276         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
289         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
302         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
315         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
328         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
341         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
361         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
374         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c0x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
387         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
400         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
413         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
426         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
439         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
452         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
465         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c2x0123, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
482         vo1x0c0123 = vmaxq_f32(vo1x0c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
492         vo1x0c0123 = vminq_f32(vo1x0c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
502           vst1q_f32(o1, vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
524             vst1q_f32(o1_tmp, vo1x0c0123); o1_tmp += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
525             vo1x0c0123 = vo1x0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
530           float32x2_t vo1x0c01 = vget_low_f32(vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
540             vo1x0c01 = vget_high_f32(vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()