94         float32x4_t vo1x0c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()  local
108         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
116         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
124         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
132         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
140         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
148         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
163         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
171         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
179         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
187         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
195         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
203         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
211         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
219         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
227         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
235         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
243         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
251         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
266         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
274         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c0x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
282         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
290         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
298         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
306         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
314         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
322         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
330         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c2x0123, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
343         vo1x0c0123 = vmaxq_f32(vo1x0c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
349         vo1x0c0123 = vminq_f32(vo1x0c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
355           vst1q_f32(o1, vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
368           float32x2_t vo1x0c01 = vget_low_f32(vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
378             vo1x0c01 = vget_high_f32(vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()