99         const float32x4_t vi1x1 = vld1q_f32(i1); i1 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
122         vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk10c0x0123, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
124         vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk10c0x4567, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
241         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk11c0x0123, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
243         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk11c0x4567, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
280         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk11c1x0123, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
282         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk11c1x4567, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
326         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk11c2x0123, vi1x1, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
328         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk11c2x4567, vi1x1, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
365         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk12c0x0123, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
367         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk12c0x4567, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
562         float32x4_t vi1x1 = vld1q_f32(i1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
588           vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk10c0x0123, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
590           vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk10c0x4567, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
719         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk11c0x0123, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
721         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk11c0x4567, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
764         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk11c1x0123, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
766         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk11c1x4567, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
821         vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk11c2x0123, vi1x1, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
823         vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk11c2x4567, vi1x1, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
856           vo0x0c0123 = vfmaq_laneq_f32(vo0x0c0123, vk12c0x0123, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
858           vo0x0c4567 = vfmaq_laneq_f32(vo0x0c4567, vk12c0x4567, vi1x1, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()