98         float32x4_t vo0x1c4567 = vo0x0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
119         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
132         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
145         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk20c0x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
165         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk00c1x4567, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
178         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c1x4567, vget_low_f32(vi1x2), 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()
204         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk00c2x4567, vget_low_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
217         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c2x4567, vget_low_f32(vi1x2), 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()
243         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c0x4567, vget_high_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
256         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_high_f32(vi1x2), 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()
282         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c1x4567, vget_high_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
295         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c1x4567, vget_high_f32(vi1x2), 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()
328         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk01c2x4567, vget_low_f32(vi0x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
341         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c2x4567, vget_low_f32(vi1x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
354         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk21c2x4567, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
367         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
380         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
393         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
406         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c1x4567, vget_high_f32(vi0x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
419         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_high_f32(vi1x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
432         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c1x4567, vget_high_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
445         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c2x4567, vget_high_f32(vi0x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
458         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
471         vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
488         vo0x1c4567 = vmaxq_f32(vo0x1c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
498         vo0x1c4567 = vminq_f32(vo0x1c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
513           vst1q_f32(o0 + 4, vo0x1c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
522             vo0x1c0123 = vo0x1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()