compute#
The compute helpers infer graph-level metadata generated from shape
inference and value lifetimes, and describe how a graph is executed: the
raw-buffer allocator, the individual execution actions
(ExecuteAction), and the ordered ExecutionPlan.
compute::ComputeContext is the single entry point that ties
these analyses together: compute::ComputeContext::Compute() runs
shape inference, value/node tagging, in-place reuse (with release-after and
shape-tag classification) and per-node peak memory in one call and keeps every
result alive; compute::ComputeContext::WriteToModel() pushes them
back into the model (value_info and node metadata); and
compute::ComputeContext::BuildExecutionPlan() derives the
ExecutionPlan from that information.