With　the　growing　demand　for　high　performance　computing,　heterogeneous　system,　based　on　CPU-GPU　architecture,　were　widely　applied　because　of　its　higher　processing　capacity.　To　reduce　the　communication　latency　between　CPUs　and　GPUs　or　other　agents,　HSA　(Heterogeneous　System　Architecture)　Foundation　proposed　a　kind　of　an　open　industry　standards　body.　In　this　paper,　we　designed　the　K-means　algorithm　on　Kaveri　APU,　which　complied　with　HSA　standard,　and　optimized　it　into　an　adaptive　one　to　obtain　high　utilization　effect　of　both　CPU　and　GPU.　Experimental　results　show　that　our　adaptive　algorithm　gets　a　45.2%　average　decrease　in　overall　execution　time.