rocm support

[Redhawk18]

Hi everyone, I'm trying to run the rocm demo and its a nightmare. I know its because of rocm and not burn, but its still a terrible experience. I've tried everything I can think of, I finally got a non-linking error and I think I'm close I'm just missing a package somewhere.

cargo bb run -b unary -B wgpu rocm --verbose

error: failed to run custom build command for `cubecl-hip-sys v6.4.4348400`

Caused by:
  process didn't exit successfully: `/home/redhawk/Desktop/school/llm/burn-bench/target/benchmarks/release/build/cubecl-hip-sys-4aa25f55ea1ed487/build-script-build` (exit status: 101)
  --- stdout
  cargo::rerun-if-changed=build.rs
  cargo::rerun-if-env-changed=ROCM_PATH
  cargo::rerun-if-env-changed=HIP_PATH
  cargo:rustc-cfg=feature="hip_42134"
  cargo::rustc-link-lib=dylib=hiprtc
  cargo::rustc-link-lib=dylib=amdhip64

  --- stderr

  thread 'main' panicked at /home/redhawk/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/cubecl-hip-sys-6.4.4348400/src/hipconfig.rs:122:5:
  Cannot retrieve the name of the HIP library directoy.
  note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
warning: build failed, waiting for other jobs to finish...
| Benchmark | Burn Version | Shapes          | Feature | Backend              | Device        | Median    |
|-----------|--------------|-----------------|---------|----------------------|---------------|-----------|
| unary-f32 | main         | (32, 512, 1024) | wgpu    | `cubecl<wgpu<wgsl>>` | DefaultDevice | 440.000µs |
| unary     | -            | -               | -       | `rocm`               | -             | FAILED    |

My two paths are set correctly, so I'm unsure whats wrong

❯ echo $ROCM_PATH
/nix/store/ssgq6f604z025kw8bnf0jys2qr8wnkym-rocm-runtime-6.3.3

burn-bench on  main [!] is 📦 v0.1.0 via 🦀 v1.89.0 via ❄  impure (devenv-shell-env) 
❯ echo $HIP_PATH
/nix/store/jv7db7h6qlggzs6adqmqq8qdwz0rpj06-hip-common-6.3.3

[laggui]

Could this be because the ROCM_PATH and HIP_CLANG_PATH differ on your machine? tracel-ai/cubecl-hip-sys#23

CC'ing @syl20bnr

[syl20bnr]

Yes, this is the same issue.

I'll take a stab at it if nobody submit a PR for it.

[Redhawk18]

Take a look on discord, trying to figure it out.

[Redhawk18]

So what variables do I need and what rocm paths do they need to be set to?

I have these three so far

              HIP_CLANG_PATH = "${pkgs.rocmPackages.hipcc}";
              HIP_PATH = "${pkgs.rocmPackages.hip-common}";
              ROCM_PATH = "${pkgs.rocmPackages.hipcc}";

[Redhawk18]

to my knowledge of looking at this code so far, it never uses env variables.

[syl20bnr]

Version 7.0.5183101 has been released which should fix the issue.