87       float32x4_t vi2x0 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
108         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
123         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
132         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
147         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
163         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
178         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
187         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
202         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
337         vi2x0 = vi2x3;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
402         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
411         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
417         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
426         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
432         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
441         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
447         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
456         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
550         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
593         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
602         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
608         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
617         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
623         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
632         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
638         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
647         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()