99         float32x4_t vo1x1c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
118         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
131         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
144         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
164         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
177         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
190         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c1x0123, vget_low_f32(vi4x2), 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()
216         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c2x0123, vget_low_f32(vi3x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
229         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c2x0123, vget_low_f32(vi4x2), 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()
255         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c0x0123, vget_high_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
268         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c0x0123, vget_high_f32(vi4x2), 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()
294         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c1x0123, vget_high_f32(vi3x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
307         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
327         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk01c2x0123, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
340         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk11c2x0123, vget_low_f32(vi3x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
353         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk21c2x0123, vget_low_f32(vi4x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
366         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c0x0123, vget_low_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
379         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c0x0123, vget_low_f32(vi3x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
392         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
405         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c1x0123, vget_high_f32(vi2x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
418         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
431         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
444         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
457         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
470         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c2x0123, vget_high_f32(vi4x3), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
487         vo1x1c0123 = vmaxq_f32(vo1x1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
497         vo1x1c0123 = vminq_f32(vo1x1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
509           vst1q_f32(o1, vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
519             vst1q_f32((float*) ((uintptr_t) o1_tmp + output_width_stride), vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
520             vo1x1c0123 = vo1x1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
532           float32x2_t vo1x1c01 = vget_low_f32(vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
535             vo1x1c01 = vget_high_f32(vo1x1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()