Remove padding from device_memory_resource

[vyasr]

Description

This PR removes the align_up call that enforces 8B padding on all RMM device memory allocations. Padding, where needed, should be the responsibility of callers, possibly with the help of a padding adapter. It should not be enforced by rmm.

I looked at all other calls to align_up in the library but didn't find any others that needed removing. The other calls all seemed to be used to guarantee specific intentional alignment requirements of other memory resources or adapters. I would appreciate verification from reviewers, though.

I've labeled this PR as breaking because it could break consumers that were implicitly relying on the current padding behavior.

Resolves #865
Resolves #1156

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[vyasr]

See #865 (comment) for more of a summary on the decision here.

[harrism]

Do you have a plan for downstream testing before merging, @vyasr ?

[vyasr]

Thanks for the reminder Mark. I've created test PRs:

• cudf: Test rmm 1278 cudf#13478
• cuml: Test rmm 1278 cuml#5448
• cugraph: Test rmm 1278 cugraph#3625
• cuspatial: Test rmm 1278 cuspatial#1165

[harrism]

I created an issue for cuSpatial as well. Assigning you @vyasr . rapidsai/cuspatial#1164

[vyasr]

Updated with a cuspatial PR.

[vyasr]

OK, we now have tests passing on all the testing PRs (cugraph's has an unrelated error blocking wheel builds). Think this is good to go now.

[vyasr]

/merge

[ttnghia]

Sorry for jumping in late. What about padding when using arena allocator? It seems that we need to modify this too:

rmm/include/rmm/mr/device/arena_memory_resource.hpp

Lines 137 to 144 in 0c08dd5

	void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
	{
	if (bytes <= 0) { return nullptr; }
	#ifdef RMM_ARENA_USE_SIZE_CLASSES
	bytes = rmm::mr::detail::arena::align_to_size_class(bytes);
	#else
	bytes = rmm::detail::align_up(bytes, rmm::detail::CUDA_ALLOCATION_ALIGNMENT);
	#endif

[jakirkham]

Could you please raise an issue?

[ttnghia]

Could you please raise an issue?

Done.