154         const float32x4_t vi2x2 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
164         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
166         vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c1x4567, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
189         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
191         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
203         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c2x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
205         vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c2x4567, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
228         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c2x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
230         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c2x4567, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
242         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c0x0123, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
244         vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk01c0x4567, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
267         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c0x0123, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
269         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c0x4567, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
281         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c1x0123, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
283         vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk01c1x4567, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
306         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
308         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c1x4567, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
400         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
402         vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
425         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
427         vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c1x4567, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
439         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
441         vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c2x4567, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
464         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c2x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
466         vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
741         const float32x2_t vi2x2 = vld1_f32(i2); i2 += 2;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
749         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c1x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
751         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c1x4567, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
764         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
766         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c1x4567, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
773         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
775         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c2x4567, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
788         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c2x0123, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
790         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
795         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()