103         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
111         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
126         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
187         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
202         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
211         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
226         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
242         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
257         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
266         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
281         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
447         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
454         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
463         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
469         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
478         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
484         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
493         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
499         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
508         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
550         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
638         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 3;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
645         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
654         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
660         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
669         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
675         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
684         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()