Skip compute-runtime fetch

[arunoruto]

I am currently trying to build llvm on NixOS, so I can package dpc++ for nixpkgs. To do so, I need to prefetch everything, and it is not allowed to fetch it otherwise by other means (due to it not being 100% reproducible). I usually fix things setup by setup, but I am stuck now at this line:

unified-runtime/cmake/FetchLevelZero.cmake

Line 111 in 6f36300

FetchContent_Populate(exp-headers)

Usually, it is possible to specify the source directory of a dependency, like UR_LEVEL_ZERO_LOADER_LIBRARY, which is an option in the same file. This isn't the case for the compute-runtime, which seems to be fetched from GitHub. Is there a possibility of skipping this using a CMake option?

My only other option is to make a patch and delete those lines, which is usually a last resort.

If someone is interested, here is my nix file for creating the derivation for building dpcpp so far.

[pbalcer]

We could just add an option to specify compute-runtime include dir. But I'm wondering, @againull, your patch added this. It looks like we use only two small structs from that entire repo. I'm wondering whether it wouldn't be better to just copy/paste it, given that it's an experimental API that will should make its way into the level-zero API if it gets stabilized.

[arunoruto]

We could just add an option to specify compute-runtime include dir. But I'm wondering, @againull, your patch added this.

I don't see a way to bypass fetching in that patch. Maybe I am missing something...?

It looks like we use only two small structs from that entire repo. I'm wondering whether it wouldn't be better to just copy/paste it, given that it's an experimental API that will should make its way into the level-zero API if it gets stabilized.

Is there a reason why the project isn't using submodules? Instead of fetching the items via git, they could be provided in the repository. The only thing that needs to be taken care of is to make sure to also fetch the submodules when cloning. Custom folders could still be provided and fetching all needed provided directories could be done with one command instead of letting it being handled by cmake files.

[againull]

We could just add an option to specify compute-runtime include dir. But I'm wondering, @againull, your patch added this. It looks like we use only two small structs from that entire repo. I'm wondering whether it wouldn't be better to just copy/paste it, given that it's an experimental API that will should make its way into the level-zero API if it gets stabilized.

We are doing sparse fetching (we don't fetch entire rpo), i.e. we currently fetch only header files with those structs and don't fetch anything else.