Add a libpressio codec wrapper

[juntyr]

• build libpressio successfully on the host
• build libpressio successfully on WASM
• add the codec wrapper

[juntyr]

@robertu94 If I compress with the noop compressor, the compressed data returned says that it has no data (but it does have the correct dtype, ndim, and number of elements). Am I doing something wrong? I do need to verify that the output has data to ensure that we're not reading out uninitialised data (UB in Rust), right?

[robertu94]

The noop compressor simply copies the input to the output. If the input has not data (e.g. pressio_data_new_empty) then the decompressed data will not either.

[juntyr]

The input does have data in my local test, and debug printing the compressed data before and after compression shows that the other metadata (dtype, dims, length) are filled in - yet the has_data call still returns false afterwards

[robertu94]

Can you please share a link to the test code, and I'll take a look (likely next week).

[juntyr]

@robertu94 Ok, I indentified the issue.

https://github.com/robertu94/libpressio/blob/868a3a70d6ebf55ad67509fbca03bdd0bc1bc246/src/plugins/compressors/noop.cc#L59-L62

While the noop compressor clones the data on compression, it copies into the decompressed data on decompression. Since I was passing in an array with no data for the decompressed output, it copied out no data, and still saw the array as empty.

Fun fact - the noop compressor is unsound (in the Rust world)! I can give it a compressed array with 5xi64 values and an empty decompressed array with shape 10xi64, and it will read 10 values from the compressed array, i.e. 5 uninitialised data values.

Is this behaviour intended? Rust cannot expose any safe code that would expose uninitialised memory, so no compressor should do this.

[robertu94]

You're right, that's a bug. What compressors are supposed to do in the decompression case is 1) if the output.has_data() == false allocate memory to hold the decompressed data. or 2) if output.has_data() == true check if output->size_in_bytes() < output_size and if so, modify the output's dimension metadata (but not capacity) to fix that.

What probably needs to happen in this case is a new function be added to the domain_manager class to handle this kind of "copy/move if it fits" mechanism so it doesn't need to be reimplmemented everywhere.

[juntyr]

Ok, now let's try to match https://github.com/robertu94/libpressio/blob/master/tools/swig/libpressio.py as much as possible ...