LazyMaterializationStrategy

A materialization strategy that defers copying of tensor view data until individual elements are actually accessed.

This strategy creates a materialized tensor that maintains a reference to the original view but caches accessed elements in a sparse data structure. Elements are copied from the view only when first accessed, providing memory efficiency for cases where only a subset of the tensor data is used.

Characteristics

• Deferred Execution: Elements are copied only when accessed

• Sparse Caching: Only accessed elements consume additional memory

• Lazy Evaluation: Computation overhead is distributed over access time

• Memory Efficient: Optimal for partial access patterns

Trade-offs

Benefits:

• Lower initial memory allocation

• Efficient for sparse access patterns

• Allows partial materialization of large views

• Maintains parent reference efficiency for unaccessed elements

Costs:

• Per-access overhead for coordinate transformation

• Additional memory overhead for caching metadata

• Potential synchronization complexity in multi-threaded access

• Parent tensor cannot be garbage collected until full materialization

Usage Scenarios

This strategy is optimal when:

• Only a subset of tensor elements will be accessed

• Memory is constrained and immediate full copying is not feasible

• Access patterns are sparse or unknown at materialization time

• The parent tensor should remain available for other operations

Parameters