Add OpenMP to radius outlier removal

[larshg]

See #6180 for reference

[larshg]

@alexeygritsenko if you can test this on your point cloud it would be nice.

[larshg]

On the test clouds, with 387 points and about 300.000 points (measured together in the test) I get something like this:
threads time(ms)
1 84
2 55
3 46
4 41
5 46

It will also depend a lot on number of neighbors.

[alexeygritsenko]

@larshg, I use C++ at a primitive level (and I don't understand anything about compilation), unfortunately I won't be able to until it becomes available through the vcpkg update command. Anyway, thanks for such a prompt improvement!

[mvieth]

Adding multi-threading to the class is a nice idea. However I would not use omp critical since these mutex locks can hinder parallelization, especially with a higher number of threads.
Here is an alternative idea: set up a std::vector<std::uint8_t> to_keep(indices_->size()) which contains a 1 if the point should be kept (enough neighbors within radius), else a 0. Fill this in parallel (no mutex lock necessary). Then, in a second for loop, transfer the results to indices and removed_indices_ (not parallel this time).

[larshg]

Adding multi-threading to the class is a nice idea. However I would not use omp critical since these mutex locks can hinder parallelization, especially with a higher number of threads. Here is an alternative idea: set up a std::vector<std::uint8_t> to_keep(indices_->size()) which contains a 1 if the point should be kept (enough neighbors within radius), else a 0. Fill this in parallel (no mutex lock necessary). Then, in a second for loop, transfer the results to indices and removed_indices_ (not parallel this time).

Yes, I reasoned that as well. I'll try your suggestion 👍

[larshg]

Looks promising - down to about 20ms now with 6+ threads. But doesn't seem to improve from there.

[mvieth]

Mostly minor things ...

Looks promising - down to about 20ms now with 6+ threads. But doesn't seem to improve from there.

I could imagine that at some point, memory access is the limiting factor instead of computing power. Either way, 20ms is really good.

[mvieth]

Why is it necessary to put this last?

[larshg]

When we copy from the original pointcloud, it also copy is_dense from the original cloud, which could be false, but after this filtering it should be true.
But I'm not sure if all filters should remove nan values if doesn't have to keep the output cloud organized.

[larshg]

Hmm, isn't the negative_ and extract_removed_indices_ kinda the same function?
Maybe the copyPointCloud (with indices) needs to update the is_dense properly, instead of just copying from the source cloud?

[larshg]

Ah, its either:
Negative_ = false, extract_removed_indices_= false -> Get those who fulfil the condition(s).
Negative_ = true, extract_removed_indices_= false -> Get those who doesn't fulfil the condition(s).
Negative_ = false, extract_removed_indices_= true -> Get those who fulfil the condition(s). removed_indices_ -> Nan and those who doesn't fulfil condition(s)
Negative_ = true, extract_removed_indices_= true -> Get those who doesn't fulfil the condition(s). removed_indices_ -> Nan and those who fulfil condition(s)

[mvieth]

Okay, so in other words: indices belonging to invalid points are always put in the removed_indices_ list (if the user asked for this list), not matter whether negative_ is true or false.

I find the logic in the final for-loop a bit too complicated, though. Currently, to_keep[i]==0 can mean to put the index into either removed_indices_ or indices, depending on the value of negative_. Similarly for to_keep[i]==1. Perhaps it would be clearer to just say to_keep[i]==0 means the index is put into removed_indices_ (if the user requested this list), and to_keep[i]==1 means the index is put into the indices list? And everything else would be handled in the parallel for-loop, including paying respect to the value of negative_. So for example, for an invalid point, to_keep[i]=0. If the point has enough neighbours and negative_==false, set to_keep[i]=1. If the point does not have enough neighbours and negative_==true, also set to_keep[i]=1. Otherwise to_keep[i]=0. Keeping the logic in the final (sequential) for-loop simple would also help parallelism.

[larshg]

Okay, so in other words: indices belonging to invalid points are always put in the removed_indices_ list (if the user asked for this list), not matter whether negative_ is true or false.

I find the logic in the final for-loop a bit too complicated, though. Currently, to_keep[i]==0 can mean to put the index into either removed_indices_ or indices, depending on the value of negative_. Similarly for to_keep[i]==1. Perhaps it would be clearer to just say to_keep[i]==0 means the index is put into removed_indices_ (if the user requested this list), and to_keep[i]==1 means the index is put into the indices list? And everything else would be handled in the parallel for-loop, including paying respect to the value of negative_. So for example, for an invalid point, to_keep[i]=0. If the point has enough neighbours and negative_==false, set to_keep[i]=1. If the point does not have enough neighbours and negative_==true, also set to_keep[i]=1. Otherwise to_keep[i]=0. Keeping the logic in the final (sequential) for-loop simple would also help parallelism.

I think I have managed to get what you suggested, though slightly different - getting inspired by the implementation in the passhrough filter.