Move cuFile linking to kvikio target

[KyleFromNVIDIA]

The exported kvikio target has an INTERFACE dependency on cuFile. Add this dependency in the local buildsystem too for consistency, and remove the usage from the examples. This will allow all targets that link against kvikio to automatically get cuFile, including the Python modules.

Fixes: #378

[bdice]

Yes, kvikio's cuFile dependency should be handled the same way we handle cuFile packages in libkvikio. This looks right to me. Thanks!

(I've only ever checked cuFile support for libkvikio because that's what we use in libcudf, so I wasn't aware of this.)

[KyleFromNVIDIA]

What I don't understand is why this doesn't already work. libkvikio has a run dependency on cuFile already:

kvikio/conda/recipes/libkvikio/meta.yaml

Lines 77 to 85 in 6ed7bcc

	run:
	- {{ pin_compatible('cuda-version', max_pin='x', min_pin='x') }}
	{% if cuda_major == "11" %}
	- cudatoolkit
	- libcufile {{ cuda11_libcufile_run_version }} # [linux64]
	- libcufile-dev {{ cuda11_libcufile_run_version }} # [linux64]
	{% else %}
	- libcufile-dev # [linux64]
	{% endif %}

[KyleFromNVIDIA]

...and it seems this change doesn't work either. I need to go back to the drawing board.

[jakirkham]

Thanks Kyle! 🙏

Think libkvikio is header only. So adding libcufile-dev doesn't solve things for other libraries, which will need to add libcufile-dev themselves

Am curious if there are more details on what is not working with this change. Would also have expected this to work

[KyleFromNVIDIA]

I'm still getting the same error:

Traceback (most recent call last):
  File "/home/kyedwards/development/kvikio/src/python/benchmarks/single-node-io.py", line 429, in <module>
    main(args)
  File "/home/kyedwards/development/kvikio/src/python/benchmarks/single-node-io.py", line 326, in main
    read, write = API[api](args)
                  ^^^^^^^^^^^^^^
  File "/home/kyedwards/development/kvikio/src/python/benchmarks/single-node-io.py", line 42, in run_cufile
    kvikio.memory_register(data)
  File "/home/kyedwards/anaconda3/envs/kvikio-env/lib/python3.11/site-packages/kvikio/__init__.py", line 10, in memory_register
    return libkvikio.memory_register(buf)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "libkvikio.pyx", line 58, in kvikio._lib.libkvikio.memory_register
RuntimeError: KvikIO not compiled with cuFile.h

[jakirkham]

Could you please share a bit more on how that is coming up? Is that one of the GHA jobs here? Or is that a local build? If so, are there any details?

---

Taking a look at one of the CUDA 12.2 x86_64 GHA jobs, it appears to be doing the right thing:

 -- Found cuFile: /opt/conda/conda-bld/_h_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeho/lib/libcufile.so
  -- Found CUDAToolkit: /opt/conda/conda-bld/_build_env/targets/x86_64-linux/include (found version "12.2.140")
  -- Found cuFile Batch API: TRUE
  -- Found cuFile Stream API: TRUE

Same story with one of the CUDA 11.8 x86_64 GHA jobs:

  -- Found cuFile: /opt/conda/conda-bld/_h_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeho/lib/libcufile.so
  -- Found CUDAToolkit: /usr/local/cuda/targets/x86_64-linux/include (found version "11.8.89")
  -- Found cuFile Batch API: TRUE
  -- Found cuFile Stream API: FALSE

Note that cuFile's Stream API was added in CUDA 12.2. So it is expected to be disabled in CUDA 11.8

Admittedly haven't checked all of the jobs. So could be missing things

[madsbk]

It looks like cmake dosn't run FINAL_CODE_BLOCK when building the Python library.

Reverting #342 fixes the issue.

[KyleFromNVIDIA]

Local testing seems to indicate that this fix works. Ready for review.

[bdice]

Wonderful! Thank you for the additional comments.

[vyasr]

Something seems off here. I agree that the changes are good for FetchContent/add_subdirectory like use cases, but IIUC we were observing issues with conda Python builds which should be finding libkvikio's config, right? Why aren't they?

[KyleFromNVIDIA]

/merge

[bdice]

Something seems off here. I agree that the changes are good for FetchContent/add_subdirectory like use cases, but IIUC we were observing issues with conda Python builds which should be finding libkvikio's config, right? Why aren't they?

This was discussed offline and is being fixed in #381.