Kernel Library
Describe tensors, validate a call, then execute numerical work against caller-owned buffers. The experimental C++20 library separates computation from application configuration, allocation services and scheduling.
| Backend | What you can use today |
|---|---|
| A65 | Twelve typed operations with checked, synchronous host execution. |
| EV74 | Native codecs for shared A65 transforms and EV-only Preproc, Histogram, FAST, FAST+BRIEF and Tracker. Compatible platform entry/transport still required; host support is not device qualification. |
Start with a working call
- Build the library.
- Call an installed kernel.
- Read tensor storage and the call lifecycle.
- Choose an operation in the support matrix.
The API is experimental (0.1.0); do not assume cross-version ABI stability.
An operation's admitted layouts, ranks and special-value rules are part of its
contract. A different operation or backend need not accept the same combinations.
Read at the right level
- Contract: observable output, numerical order, supported tensors and errors.
- Implementation notes: how the current release selects SIMD, packing and fallback paths.
- API reference: declarations generated from public headers.
- Contributor guide: independent tests, measurements and documentation.
Implementation names and dispatch thresholds explain this source revision; they are not an application-level placement policy or a permanent API guarantee.