Expand description
ROCm 6.x FFI + dynamic loader — mirror of iron_cuda_sys for AMD GPUs.
Every library is dynamically loaded on first use. Nothing is linked at build time, so the workspace compiles on any machine with or without ROCm installed.
Supported libraries:
hip— HIP runtime (driver + runtime merged, as ROCm ships it)hipblas— hipBLAS dense linear algebra (Level 1/2/3 + batched)hipblaslt— hipBLASLt matmul front-end (FP8 on CDNA3+)hiprtc— HIPRTC runtime kernel compile (the NVRTC analogue)rccl— AMD collective comms
Everything else (rocFFT, rocSPARSE, rocSOLVER, MIOpen, rocRAND) will
be added as needed. Pattern is identical: add a module, implement
load_fns, stash the result in a LazyLock<Result<Fns, LoadError>>.
§Documentation policy
Items here mirror the ROCm C headers one-to-one (each Fns field is
named after the C symbol it resolves), so they are not documented item
by item — AMD’s reference for the same symbol is the authority. The safe
crate (ironaccelerator-rocm) is fully documented and enforces
missing_docs.
Modules§
- hip
- HIP runtime FFI. ROCm ships HIP as a single library (
libamdhip64) that unifies driver- and runtime-scoped entry points, unlike CUDA’s split. - hipblas
- hipBLAS — dense Level 3 matmul. We bind the legacy handle API because
that’s what’s stable across rocm 6.x; hipBLASLt (the FP8-era front-end)
lives in
crate::hipblaslt. - hipblaslt
- hipBLASLt — the Lt front-end used for FP8 matmul on CDNA3+ (
gfx942). - hiprtc
- HIPRTC — runtime compilation of HIP C++ to a device code object.
- loader
- Dynamic library loader. Identical contract to the CUDA sys crate.
- rccl
- RCCL — AMD collective comms. API parity with NCCL.