144         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()  local
151         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
160         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
166         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
175         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
181         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
190         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
196         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
205         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
247         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
336         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 3;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()  local
343         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
352         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
358         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
367         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
373         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
382         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()