CI: add hip quality check

[IMbackK]

Add "hip-check" ci workflow, this workflow is intended to solve the following problems for us:

1. The hip backend has had repeated problems with people adding #pragma unroll loops that cant be unrolled see ssm-conv, softmax and CUDA: use shared mem for ssm_conv #20128 (comment)
   • this workflow builds the hip backend with Werror to make it obvious when these occure
   • i did not make the hip builds Werror generally as we can not guarentee that there will be no warnings when built against every rocm version we support
2. Currently we want to support rocm >=6.1 so build for 6.1 in the build workflow to check that no functions from newer rocm sneak in, however we build the release against the latest version of rocm, this causes the issue that we dont notice if a pr dosent build against the latest version until the release fails
   • This workflow builds against the same version of rocm as the release to check for this
3. GCN/CDNA have a rather small 64 KiB vector register file and are wave 64, which means that only 256 registers are available. Quite often kernels get added that include significant register spill.
   • This pr adds a small script that is used in the workflow to check for kernels that spill significant registers
   • The script ignores kernels with significant spills currently in the code base (altho several of these should be investigated at some point)

The idea of this workflow that its failure NOT necessarily block a pr, but that its should prompt investigation / possible adding to the workflows witelist.

[IMbackK]

theres a run here https://github.com/IMbackK/llama.cpp/actions/runs/22975551980/job/66703103556 to show how this works (the failure in the VGPR check is on purpose for illustration)

[IMbackK]

@CISC the hip compiler will ignore #pragma clang diagnostic push when --save-temps is active. For this reason a build with -DGGML_HIP_EXPORT_METRICS=On and -Werror can not succeed as warning will be emmited here:

llama.cpp/ggml/src/ggml-cuda/softmax.cu

Line 51 in f2ab047

#pragma clang diagnostic push

among other places.

[CISC]

Better split the jobs again then.

[IMbackK]

Right

[CISC]

Check CI failures.

[IMbackK]

this pr should be in principal ready now, but i think this needs some buy in from the maintainers of the cuda backend this also affects.

@JohannesGaessler @am17an @slaren thoughts on principal?

[CISC]

Guess we got one:
https://github.com/ggml-org/llama.cpp/actions/runs/23210660849/job/67458238374?pr=20430#step:6:155

[IMbackK]

Guess we got one: https://github.com/ggml-org/llama.cpp/actions/runs/23210660849/job/67458238374?pr=20430#step:6:155

yeah i opened a pr for that: #20696

[am17an]

Can we store the unmangled names? Are these not compiler dependent

[IMbackK]

Name mangling is in theory compiler dependent, however on linux all compilers follow the itanium abi spec for mangling so this is not an issue for this workflow. Additionally, HIP code can only be compiled by clang anyhow and this script depends on a specific feature of the amdclang fork (the metrics output).

The compiler dose not output demangled names in its hip metrics output so the only way to have unmangled names here would be for the script to unmangle the names itself in accordance to itanium spec im not sure this is worth it.

[JohannesGaessler]

Based on static code analysis this seems correct to me.