PMF Ground Classification Multithreading?

[myerbro3]

Hello all,

I am running a ground classification with the PMF algorithm.

library ("lidR")

is.parallelised(pmf(ws = 5, th = 3))
# FALSE
lasclass <- classify_ground(las, algorithm = pmf(ws = 5, th = 3)) 
# Single Thread

A colleague tells me she has been able to run this on multiple threads but I cannot replicate her results, something like:

Morphological filter: 1% (8 threads)

From the docs, I can use is.parallelised() and see that the PMF algorithm is not parallelised, confirmed by running the classification I see:

Morphological filter: 1% (1 threads)

Our script is the same. I have used set_lidr_threads() for which get_lidr_threads() returns 32. Could she be mistaken, or am I missing something here?

Thank you, looking forward to any advice.
-Brooks

[Jean-Romain]

Good question. According to the source code it is not parallelized so is.parallelised() is correct. Why it is not parallelized? I don't know. It could be parallelized, I don't see anything that make it hard to parallelize.

About your colleague, I don't know. I swear there is no parallelization directive in the code.

[myerbro3]

Jean-Romain,

Thank you for your quick response. Is parallelizing this code something that I could do as a user? I am proficient in R but have no experience with writing or editing packages.

Alternatively, I see in the docs that parallelization is made possible through a LAScatalog using something like this from 14.7.4:

library(future)
plan(multisession)
opt_chunk_size(ctg) <- 400
opt_chunk_buffer(ctg) <- 40
opt_output_files(ctg) <- paste0(tempdir(), "/{*}_classified")
classified_ctg <- classify_ground(ctg, csf())

If I were to use the above code, replacing csf() with pmf(), should I see an increase in performance compared to processing each .las individually? I have ~280 GB of high resolution UAS .las data (roughly 8 billion points), currently in 113 ~1-3 GB tiles, but these can be re-tiled to whatever size I need. I have an i9-13900K and 64 GB RAM.

Best,
Brooks

[Jean-Romain]

Is parallelizing this code something that I could do as a user?

It is probably a single line of code to change in the c++ source code here

lidR/src/LAS.cpp

Line 142 in 5a5a3a5

void LAS::z_open(double resolution)

But we must ensure it is really parallelizable first.

If I were to use the above code, replacing csf() with pmf(), should I see an increase in performance compared to processing each .las individually?

You will read, classify and write multiple files simultaneously. It will very likely be more performant if done properly; don't use all cores, ensure your computer has enough memory, stuff like that. 64 GB is a lot but lidR is far from optimal in terms of memory usage by design. If you put 6 or 8 cores on it you may compute 4 times faster. As always with parallelization, testing it first is the only solution to measure the actual gain.

[myerbro3]

I see, thank you for this very helpful information. I'll give the LAScatalog a shot and see how it goes.

[Jean-Romain]

plan(multisession)

will use all your cores. Use

plan(multisession(workers = 6))