Cuda NMS argsorts with invalid entries

I believe the following two lines should have valid_count=valid_count, as in the default implementation. I believe the data array is zeroed at initialization for cuda, so the issue only manifests if nms is run more than once, with the first run containing more valid boxes than the second run.

https://github.com/apache/incubator-tvm/blob/21002cd094c34716b1e07a63ed76f53dadd60e23/python/tvm/topi/cuda/nms.py#L488

https://github.com/apache/incubator-tvm/blob/21002cd094c34716b1e07a63ed76f53dadd60e23/python/tvm/topi/cuda/nms.py#L492

[zhiics]

cc @Laurawly

[Laurawly]

@alec-anyvision Thanks for pointing this issue out. The changes made were explained in PR #5339. So in the original implementation, get_valid_count will rearrange the invalid boxes to the end of the array while in the updated implementation, this movement is done by argsort to improve the performance. valid_count is changed to initialized to None in order to let argsort sort the entire array instead of only the first valid_count number of elements. Since argsort doesn't make changes to valid_count, I'm not sure why running nms multiple times will manifest the number of valid boxes. Could you elaborate more?

[masahi]

This code is no longer used.