142         const float32x4_t vi2x2 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
149         vo1x1 = vfmaq_laneq_f32(vo1x1, vk00c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
162         vo0x1 = vfmaq_laneq_f32(vo0x1, vk20c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
170         vo1x1 = vfmaq_laneq_f32(vo1x1, vk00c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
183         vo0x1 = vfmaq_laneq_f32(vo0x1, vk20c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
191         vo1x1 = vfmaq_laneq_f32(vo1x1, vk01c0, vi2x2, 2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
204         vo0x1 = vfmaq_laneq_f32(vo0x1, vk21c0, vi2x2, 2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
212         vo1x1 = vfmaq_laneq_f32(vo1x1, vk01c1, vi2x2, 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
225         vo0x1 = vfmaq_laneq_f32(vo0x1, vk21c1, vi2x2, 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
280         vo1x0 = vfmaq_laneq_f32(vo1x0, vk02c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
293         vo0x0 = vfmaq_laneq_f32(vo0x0, vk22c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
301         vo1x0 = vfmaq_laneq_f32(vo1x0, vk02c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
314         vo0x0 = vfmaq_laneq_f32(vo0x0, vk22c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
396         float32x4_t vi2x2 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
403           vi2x2 = vld1q_f32(i2 + 4);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
411         vo1x1 = vfmaq_laneq_f32(vo1x1, vk00c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
424         vo0x1 = vfmaq_laneq_f32(vo0x1, vk20c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
432         vo1x1 = vfmaq_laneq_f32(vo1x1, vk00c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
445         vo0x1 = vfmaq_laneq_f32(vo0x1, vk20c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
454           vo1x1 = vfmaq_laneq_f32(vo1x1, vk01c0, vi2x2, 2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
471           vo0x1 = vfmaq_laneq_f32(vo0x1, vk21c0, vi2x2, 2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
481           vo1x1 = vfmaq_laneq_f32(vo1x1, vk01c1, vi2x2, 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
498           vo0x1 = vfmaq_laneq_f32(vo0x1, vk21c1, vi2x2, 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
556           vo1x0 = vfmaq_laneq_f32(vo1x0, vk02c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
565           vo0x0 = vfmaq_laneq_f32(vo0x0, vk22c1, vi2x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
571           vo1x0 = vfmaq_laneq_f32(vo1x0, vk02c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
580           vo0x0 = vfmaq_laneq_f32(vo0x0, vk22c2, vi2x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()