Design notes¶
The reasoning behind particular choices in the schemas, for readers who want to know why a document has the shape it has. Each note stands alone.
- Why a model is a graph
- Why a model is described as the GPU work it launches rather than in its vendor's configuration format, and why the derived graph is committed.
- Where the kernel stops
- How the
Kernelinterface divides work between a cost model and a simulator, and why a request and its resolution are kept apart. - Parallel layouts
- How a pool's widths determine its rank count and expert group, and the checks that decide whether a layout could be launched.
For more detail than these notes give, read the source. Each package begins with a comment that says what it is for and what it leaves out.