[UR] Enable hwloc in UMF linux build

[PatKamin]

Update UR to statically link UMF with hwloc on Linux.
This will enable hwloc support in UMF on Linux.

Also, always use the v2.9.3 hwloc version for hwloc builds.

[PatKamin]

Drafted until the UMF fix is pulled down to llvm via UMF update in UR.

[PatKamin]

Drafted until the UMF fix is pulled down to llvm via UMF update in UR.

Fix is now pulled down, ready for review.

[igchor]

You should probably also update https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedGuide.md (remove hwloc as a dependency that needs to be installed by the user)

[PatKamin]

You should probably also update https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedGuide.md (remove hwloc as a dependency that needs to be installed by the user)

Right, now it will automatically be downloaded - deleted user requirement.

[bader]

Update UR to statically link hwloc v2.9.3 both on linux and Windows.

If I get the change right, the dependency on hwloc still exists. The change removed the runtime dependency, but not the compile time dependency. Am I right?

This document describes the environment for DPC++ compiler developers, not DPC++ compiler users.

Your change seems to raise the version requirement. Please, update the hwloc version to 2.9.3.

[PatKamin]

The hwloc library will always be fetched from sources. Regardless of hwloc-dev packages user has installed, the fetched version will be linked with the UMF library. That's why I remove hwloc from this dependencies list, as the user-installed hwloc-dev package will not be used by UMF.

[bader]

You should probably also update https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedGuide.md (remove hwloc as a dependency that needs to be installed by the user)

Right, now it will automatically be downloaded - deleted user requirement.

If hwloc dependencies will be automatically resolved by CMake, please, remove installation of these dependencies from CI scripts. Specifically, from

llvm/devops/scripts/install_build_tools.sh

Line 26 in 7cc9e80

libhwloc-dev \

and

llvm/.github/workflows/sycl-docs.yml

Line 36 in 7cc9e80

sudo apt-get install -y graphviz ssh ninja-build libhwloc-dev

.

[PatKamin]

In general, I'm not a fan of having those dependencies because they complicate our SW stack for not so apparent reason. Pardon my ignorance, but why hwloc and even UMF have to be used by default? At least for the amount of testing we perform at intel/llvm - are there any tests which can detect side effects of UMF and/or hwloc being present or absent? If not, then I would be strongly against bringing that dependency by default, because it means that its an overhead without a reason.

The UMF library is required for allocation handling by default in UR, the user cannot opt-out, UR can't be built without UMF. There are two ways that UR gets UMF now. UMF is either fetched by UR or the user can force CMake to search for the installed UMF with find_package using the UR_USE_EXTERNAL_UMF env var.
Although currently hwloc can be disabled and not used, the env var that lets the user to do this (UMF_DISABLE_HWLOC) is planned to be removed prior the stable 1.0 release of UMF. hwloc will be a required dependency of UMF.

Also, why do we need to build either from source at all? I understand reason for building UR from sources, because it is moves together with SYCL RT on a very fast pace and changes have to be done in sync, but is it really the case for UMF and hwloc?

We don't release UMF prebuilt libraries as a part of open-source releases and there are no plans to do so. Users using UMF have to fetch it.

As @bader suggested, why can't we just download a binary release of hwloc (and ideally UMF) during UMF (UR) build?

I suggest we fetch binaries instead of sources.
https://www.open-mpi.org/software/hwloc/v2.9/ has links to the binaries for different platforms.

I don't fetch hwloc binaries as we apply a patch to its sources.
I also wanted the intel/llvm hwloc handling to be the same as on the downstream where I wouldn't like to force having hwloc binaries installed in the system as a requirement.

On the second hand, currently on Linux hwloc is a requirement of intel/llvm and CI seems to be ready for enabling hwloc in UMF Linux builds, I could verify that. So I propose to stick with having hwloc as a requirement in intel/llvm and not fetch it.

[vinser52]

AS @PatKamin said today UMF project does not distribute binaries. We can prioritize this discussion and estimate what should be done to distribute UMF in binary form. I agree that it might be a best option.

Regarding why UMF is not optional, @PatKamin said that UR uses UMF to support memory pooling. Another reason is the interop between runtimes. UMF provides observability for other runtimes. For example, MPI uses UMF to get the memory properties of the memory allocated by SYCL applications.

[AlexeySachkov]

I don't fetch hwloc binaries as we apply a patch to its sources.

Does it mean that we need to redistribute our own version of hwloc then?

[PatKamin]

Does it mean that we need to redistribute our own version of hwloc then?

No, it's just one of the options. The other is to require the user to have hwloc installed.

[PatKamin]

Reverted back changes with fetching hwloc. Leaving hwloc as a required dependency. Failing CI jobs also fail on other PRs, looks like these failures are not caused by changes in this PR.

[AlexeySachkov]

Thanks for educating me on the topic, @vinser52, @PatKamin.

I don't fetch hwloc binaries as we apply a patch to its sources.

Does it mean that we need to redistribute our own version of hwloc then?

No, it's just one of the options. The other is to require the user to have hwloc installed.

What is the point of patching hwloc if we link with it dynamically and don't redistribute the updated version? Shouldn't we link statically with hwloc then?

Do I understand correctly that we don't fetch hwloc on Windows either and it is also expected to be present on a build system?

[PatKamin]

What is the point of patching hwloc if we link with it dynamically and don't redistribute the updated version? Shouldn't we link statically with hwloc then?

The patch is needed to satisfy Coverity requirements in case hwloc is fetched and built as a static library.

Do I understand correctly that we don't fetch hwloc on Windows either and it is also expected to be present on a build system?

In Windows we have to do the fetch and link statically with hwloc always: https://github.com/intel/llvm/blob/sycl/sycl/cmake/modules/FetchUnifiedRuntime.cmake#L129-L130

Please, note that I've minimized the changes in this PR recently. On current sycl branch, hwloc already is a required llvm dependency and is installed in intel/llvm CI environment. I've removed the fetching for Linux builds in favor of using the installed hwloc library to reduce the unnecessary overhead of hwloc fetch and build during llvm build.

This will, however, require changes in the downstream CI environment, so I put this PR on draft until it's ready there.

[PatKamin]

Changes wait for #17551 to be merged.

[PatKamin]

@intel/unified-runtime-reviewers, please take a look - all tests passed. Changes are now minimal.

[PatKamin]

@intel/llvm-gatekeepers, I think this PR is ready to merge

[ldrumm]

It doesn't look like Alexey's concerns have been addressed so I'm going to hold off merge until he's confirmed he's okay with this.

As a side note: the PR description doesn't do what it says

[maksimsab]

@PatKamin

I am getting the following error locally. Can it be related to this change?

[1231/8769] Generating ../hwloc_targ-src/configure
FAILED: _deps/hwloc_targ-src/configure /localdisk2/msabiani/llvm/build/_deps/hwloc_targ-src/configure
cd /localdisk2/msabiani/llvm/build/_deps/hwloc_targ-src && ./autogen.sh
./autogen.sh: line 13: autoreconf: command not found

[PatKamin]

@PatKamin

I am getting the following error locally. Can it be related to this change?

[1231/8769] Generating ../hwloc_targ-src/configure
FAILED: _deps/hwloc_targ-src/configure /localdisk2/msabiani/llvm/build/_deps/hwloc_targ-src/configure
cd /localdisk2/msabiani/llvm/build/_deps/hwloc_targ-src && ./autogen.sh
./autogen.sh: line 13: autoreconf: command not found

The hwloc is required on Linux (

llvm/sycl/doc/GetStartedGuide.md

Lines 47 to 48 in 0886791

	* `hwloc` version 2.3 or later (Linux only)
	* libhwloc-dev or hwloc-devel package on linux

), otherwise fetching triggers.