Expand description
§ironaccelerator-rocm
ROCm / HIP backend. Targets the ROCm 6.x API surface and the AMD datacenter accelerator stack:
- HIP driver + runtime (analogous to CUDA driver/runtime)
- hipBLAS / hipBLASLt — dense matmul, FP8 on CDNA3+
- MIOpen — convolution + fused kernels (cuDNN analogue)
- rocFFT, rocRAND, rocSPARSE, rocSOLVER
- RCCL — collective comms (NCCL analogue)
- Composable Kernel — tunable templates (CUTLASS analogue)
§Status
Safe HIP driver wrappers (drv), HIPRTC kernel compilation with a
race-free in-memory module cache (kernel), a size-class memory pool
(pool), and a hipBLASLt matmul surface (blas). Device paths compile
but are not yet validated on AMD hardware.
Re-exports§
pub use backend::RocmBackend;pub use backend::ROCM_BACKEND;
Modules§
- backend
- ROCm
Backendimpl. Live device enumeration through HIP. - blas
- hipBLASLt safe wrapper — mirrors
ironaccelerator_cuda::blas. - drv
- Safe HIP driver layer. Parity with
ironaccelerator_cuda::drv, minus the primary-context dance: HIP manages a per-thread current device implicitly. - kernel
- HIPRTC runtime kernel compilation + an in-memory module cache — the ROCm
analogue of
ironaccelerator_cuda::kernel. - pool
- Per-stream small-buffer pool for ROCm — the AMD analogue of
ironaccelerator_cuda::pool::MemPool.
Functions§
- register
- Registers the ROCm backend with
reg.