Expand description
Optional runtime-loaded ROCm/rocBLAS tensor compute provider.
The ROCm provider has no link-time ROCm dependency. It validates HIP and
rocBLAS entry points plus an observed AMD gfx* target at runtime and
exports a tensor site only after that ABI evidence exists. The executor
accepts dense matmul for f32 plus half-family dtypes when rocBLASLt evidence
is present.
Structs§
- Compute
Rocm Lib - Loadable library that registers a ROCm compute site when runtime validation succeeds.
- Fake
Rocm Loader - Fake loader for deterministic tests.
- Rocm
AbiEvidence - Dynamic-library ABI evidence required by the ROCm provider.
- Rocm
Allocation - Opaque ROCm allocation evidence.
- Rocm
Library Set - Loaded ROCm runtime libraries kept alive for function-pointer validity.
- Rocm
Load Error - ROCm dynamic-loading failure.
- Rocm
Resident Storage - Resident storage for a ROCm matmul result.
- Rocm
Runtime Loader - Real dynamic loader using platform ROCm shared libraries.
- Rocm
Runtime Probe - Result of ROCm runtime discovery.
- Rocm
Symbol Evidence - One validated runtime symbol.
- Rocm
Tensor Executor - Tensor executor backed by validated ROCm/rocBLAS ABI evidence.
Statics§
- RECIPES
- Cookbook recipes for this lib, embedded at build time.
Traits§
- Dynamic
Rocm Loader - Loader abstraction used by real and fake ROCm discovery.
Functions§
- compute_
rocm_ capability - Capability required before a ROCm hardware tensor site can be realized.
- compute_
rocm_ lib_ symbol - Stable symbol for the ROCm runtime library.
- compute_
rocm_ site_ symbol - Site symbol exported by a successful ROCm runtime.
- discover_
rocm_ runtime - Discovers ROCm using the real platform dynamic loader.
- rocm_
executor_ symbol - Stable symbol for the ROCm tensor executor.