CUB::BlockRadixSort: how to deal with the last tile which is not full?

问题

There are 510 keys for sort. BLOCK_DIM_X = 128, ITEMS_PER_THREAD = 4, thus every tile covers 512 keys. We lauch kenel by 1 block.

my kernel looks like this:

    typedef cub::BlockRadixSort<int, 128, 4>  BlockRadixSort;
    int thread_data[4];
    BlockLoad(temp_storage.load).Load(in_data, thread_data);
    CTA_SYNC();

    BlockRadixSort(temp_storage.sort).Sort(thread_data);
    CTA_SYNC();

    BlockStore(temp_storage.store).Store(out_data, thread_data);
    CTA_SYNC();

The problem is BlockRadixSort sort 512 keys, not 510. How to exclude the last 2 items from block sort?

来源：https://stackoverflow.com/questions/62170084/cubblockradixsort-how-to-deal-with-the-last-tile-which-is-not-full