296         const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()  local
299         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
300         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
304         vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk21c1x0123, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
305         vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk21c1x0123, vget_high_f32(vi4x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
682         const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()  local
685         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
686         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
970         const float32x4_t vk21c1x0123 = vld1q_f32(w + 120);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()  local
973         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk21c1x0123, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
974         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()