Basic concepts

The term provenance refers to the matching of device code to framework subgraphs; it is analogous to source code locators in conventional compilers, which associate regions of object code with source files and line numbers. Provenance is extensible in that it may also include the chain of passes that lead from the framework graph to the executing code.

It can associate device code with specific tags added by a framework bridge which correspond to the framework ops that create the nGraph nodes. This works only for those transformations that take place in nGraph: the information stored in the nodes can include additional details about how the device code was chosen. For example, whenever a graph transformation is performed with one of the nGraph core Ops, a lower level of abstraction can record information about the transformation that may be useful to anyone wondering why a kernel was “chosen”; a complete description of the steps leading to the device kernels being used, as well as all of the framework nodes that led to the kernel, can be obtained.

Existing use cases

Currently, every node nGraph touches can optionally have a set of provenance tags, which are strings set by a framework bridge. When a set of nodes is replaced by a new set of nodes, a combination of heuristics and special casing is used to set the tags on the new nodes based on the tags from the old nodes.

A builder is a function that creates a sub-graph and returns a root node to the bridge. The bridge is not necessarily aware of the subgraph, only of the returned node, where it sets tags. The remaining nodes’ tags are set by associating a set of nodes, called a provenance group, with the node. Any tags added to the node are also added to the nodes in the provenance group.

An updated implementation of the functionality of builders is the fused op, a node that can replace itself with a subgraph. When the node is expanded into a subgraph, a vector of values is returned, corresponding to outputs of the original fused op; the tags of the fused op are added to all nodes in the values in reverse dataflow direction, up to (though not including) the input values of the fused op.