Infer.NET development : Compiler transforms
Message optimisation transform
Removes duplicate messages or redundant message operations in order to reduce the memory/computation needed for inference.
| Optimisation | Description | Action | Saving |
|---|---|---|---|
| Collapse identical composite message arrays | Composite message arrays are unrolled and a ResultIndex supplied to the operator. For some operators e.g. VMP UsesEqualsDef, the returned message does not depend on the result index i.e. all messages are the same. Note: no need for new annotation |
THIS WILL BE DONE BY THE BELOW IF WE USE THE BUFFER TRANSFORM. We can remove duplicates by collapsing the message array - make the array length 1 and setting all indices of the array to 0. Arrays to be collapsed could be marked in MessageTransform, but the collapsing would be achieved here. |
Memory saving of ~50% for VMP algorithms. |
| Remove redundant operations | Some operators (e.g. copy operators) return one of their arguments directly as the return value. These arguments should be annotated with [IsReturned]. | Any reference to the result of such an operator should be replaced with the argument expression instead (copying across dependencies as appropriate). The operators could be marked in MessageTransform and the replacement done here. |
Memory overhead and small computation overhead of copy operations. Cleaner generated code. |
| Collapse unit length arrays | to be done? makes most sense for composite arrays e.g. uses | - | Small memory savings, but mainly cleaner generated code. |