Matrix Decimation Streaming Engine for Strided Convolutions

The streaming engine is divided into multiple functional units including an address generation unit, a data formatting unit, and a decimation unit. Each unit handles a specific aspect of data stream processing, allowing the complex function to be broken down into manageable segments that can operate independently and in parallel, thus improving overall productivity without proportionally increasing complexity.

The streaming engine implements six levels of nested iteration support, where inner iteration loops are contained within outer iteration loops. This nested structure allows complex data patterns to be processed through hierarchical loops, with each level managing a specific dimension of data access, enabling efficient handling of multi-dimensional data streams while maintaining structured control flow.

The streaming engine performs preliminary formatting and separation of matrix vector elements before they are stored in the register file. The address generation unit pre-calculates memory addresses and the formatting unit pre-organizes data elements according to the required access pattern. This preliminary processing ensures that when data is scheduled for processing, it is already in the optimal configuration, improving data scheduling efficiency without requiring additional processing time during the main computation.

A buffer memory is introduced as an intermediary between the memory controller and the register file. The streaming engine writes formatted and separated vector elements to this buffer, which then supplies data to the register file at the required rate. This intermediary structure decouples the writing and reading operations, allowing data separation and formatting to occur without directly blocking the processing timeline.