Add gdeflate compression and decompression with NVIDIA fork of libdeflate.

[MarkLeone]

• Gdeflate compression can yield multiple 64K pages.
• Each chunk includes number of pages and size of each page.
• Off by default. gdeflate calls are wrapped and conditionally compiled in.
• CMake and Bazel builds can optionally fetch NVIDIA fork of libdeflate.
  Additional info and discussion on openexr-dev@lists.aswf.io

[linux-foundation-easycla]

• ❌ - login: @MarkLeone / name: Mark Leone . The commit (12fe744) is not authorized under a signed CLA. Please click here to be authorized. For further assistance with EasyCLA, please submit a support request ticket.

[cary-ilm]

Thanks for this! I'll look into the failing CI checks this weekend, I think I know what's going on.

The project policy is to require a signed Contributor License Agreement. I don't recall if Nvidia has an agreement in place, but is that something you can follow up on?

Also, we require signed git commits. Could you amend your commit with "-s"? That adds the "Signed-off-by" line. Note it's not sufficient to submit an additional commit, they all have to be signed, so you'll need to amend the commit and force-push. Or close this PR and submit a fresh one.

[kmilos]

NOT OPENEXR_FORCE_INTERNAL_DEFLATE AND OPENEXR_FORCE_FETCHCONTENT_DEFLATE?

Otherwise the check for gdeflate support below doesn't work if both are somehow set to ON. Or have a mechanism (throw an error?) to ensure they're exclusive?

[Vertexwahn]

we could have a flag - like in https://github.com/bazelbuild/bazel-central-registry/blob/0040dfca0c14b4eda9e079fb51fde2567c39a8d9/modules/boost.asio/1.89.0/overlay/BUILD.bazel#L5

In boost.asio one can decide to use OpenSSLvs. BoringSLL vs. No SSL support

Untested (inspiered by boost.asio):

string_flag(
    name = "defalte",
    build_setting_default = "libdeflate",
    values = [
        "libdefalte",
        "ibdefalate_nvl",
    ],
    visibility = ["//visibility:public"],
)

config_setting(
    name = "libdeflate",
    flag_values = {":defalte": "libdeflate"},
)

config_setting(
    name = "libdeflate_nv",
    flag_values = {":defalte": "libdeflate_nv"},
)

...

    deps = [
      ...
    ] + select({
        ":libdeflate": ["@libdeflate//:deflate"], # or maybe use default attirbute here
        ":libdeflate_nv": ["@libdeflate_nv//:deflate"]
    }),

Someone who uses OpenEXR (or your CI job) can than add the flag
build --@openexr//:deflate=libdeflate_nv (this could also be added to the .bazelrc)

[cary-ilm]

@MarkLeone, looking at this a little deeper, I think the better configuration is to simply vendor in the source code directly from NVIDIA/libdeflate, and avoid the FetchContent altogether. Look in share/util/check_deflate.sh, which is a helper script to be run manually whenever a new version of libdeflate is released. In essence, that's what you've done, but with a forked repo. I think if you change the URL in that script to https://github.com/NVIDIA/libdeflate, then run it, the CMake configuration doesn't to change much at all.

src/lib/OpenEXRCore/compression.c includes only the files out of libdeflate that are needed, the rest are ignored. You'll need to add whatever gdeflate-related files are needed there.

[cary-ilm]

@MarkLeone, it looks like NVIDIA/libdeflate is based on a very old fork of ebiggers/libdeflate (v1.8 from 2021), which was before libdeflate introduced cmake support, which complicates the setup, since your library has no cmake config. I started working on fixing the CI to test against an external build of NVIDIA/libdeflate, but that's going to require some extra hoops to jump through.