97         float32x4_t vo0x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
117         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
130         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 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()
156         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c1x0123, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
169         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 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()
202         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c2x0123, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
215         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c2x0123, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
228         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c2x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
241         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c0x0123, vget_low_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
254         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c0x0123, vget_low_f32(vi1x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
267         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c0x0123, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
280         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c1x0123, vget_high_f32(vi0x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
293         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c1x0123, vget_high_f32(vi1x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
306         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
319         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk01c2x0123, vget_high_f32(vi0x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
332         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk11c2x0123, vget_high_f32(vi1x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
345         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk21c2x0123, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
365         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c0x0123, vget_low_f32(vi0x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
378         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c0x0123, vget_low_f32(vi1x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
391         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
404         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c1x0123, vget_low_f32(vi0x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
417         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
430         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
443         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
456         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
469         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c2x0123, vget_high_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
486         vo0x1c0123 = vmaxq_f32(vo0x1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
496         vo0x1c0123 = vminq_f32(vo0x1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
512           vst1q_f32(o0, vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
521             vst1q_f32((float*) ((uintptr_t) o0_tmp + output_width_stride), vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
522             vo0x1c0123 = vo0x1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
531           float32x2_t vo0x1c01 = vget_low_f32(vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
537             vo0x1c01 = vget_high_f32(vo0x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()