108         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
118         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
120         vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
143         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
145         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
157         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
159         vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
182         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
184         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
276         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
278         vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
301         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
303         vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
315         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
317         vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk01c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
340         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
342         vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
361         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
363         vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
386         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
388         vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
400         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
402         vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
425         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
427         vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
665         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
670         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
672         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
685         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
687         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
694         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
696         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
709         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
711         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
718         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
720         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
733         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
735         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
742         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
744         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
757         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
759         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
795         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
951         const float32x2_t vi2x1 = vld1_f32(i2); i2 += 2;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
956         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
958         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
971         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
973         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
980         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
982         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c2x4567, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
995         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
997         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()