Document mgard operator

[eisenhauer]

The mgard compression operator should be documented along with the others (I think it was merged experimentally, but regardless, it deserves docs if it's on the source tree). Additionally, the remaining lossless_types should be configurable.

[anagainaru]

When documenting the compression operators, I would like to have a field on how they deal with GPU pointers (namely, can they work with input GPU and output CPU, which is required by the GPU backend).

[eisenhauer]

Hmm. How should this be working? Our Operator class doesn't carry the MemorySpace parameter as an input. Don't they all operate on CPU at present? Copying from GPU to CPU space handled outside of the operator?

[anagainaru]

Some of the operators detect where the buffer is, so when you give adios a GPU pointer it just passes it to the operator. I cannot control the output buffer provided because we decided not to allocate memory on the GPU so it is a CPU pointer. The operator detects this and does a member copy after compression. Similar for decompression but in the opposite order. We tested this with a bunch of compressions (I think Blosc and ZFP) and they were working. I think MGARD can do this too (we just need to document which ones can handle it).

[eisenhauer]

We maybe want to rationalize our MGARD compressor situation before we document it like this. We likely don't need three mgard operators.

[eisenhauer]

So, I did take a quick look at this (at least the possibility of mgard merge). Obviously having a different operator for different types isn't consistent with the rest of ADIOS, but merging is hard. MGardComplex allocates a new block of host memory on every operate/inverse to hold the split arrays. I guess the operator can copy into and out of GPU memory and run the middle bits on the GPU, but it's not going to behave the same as with simple float/double, etc. I think I'll let this lie and let more knowledgeable folks tackle it...

[pnorbert]

@gqian-coder should document how the complex version of mgard works.

MGARD handles GPU/CPU memory pointers by itself and will run on GPU/CPU accordingly. ZFP and Blosc are either GPU or CPU only (libraries), so their operator code has to take care for downloading/uploading data.