How yonseicasl/neurospector works
Main Components
• Command-Line Interface & Main Entry: Represent the main executable (src/main.cc) that interprets runtime options.
• Mode Switch: A branch that directs execution into either the “optimizer” module or the “analyzer” module.
• Optimizer Module: Consisting of multiple strategies (see src/optimizer with bottom-up and brute-force implementations plus mapping space definitions). This module takes accelerator and network configurations to compute optimal dataflows and mappings.
• Analyzer Module: Represented by src/analyzer along with common modules to compute cost metrics based on a provided scheduling table.
• Common Libraries/Utilities: Components such as the “common” folder (handling accelerator, network, layer, and scheduling table objects) and “utility” (including the parser, version info, and other helper functions).
• Configuration Files: The different configuration directories under “configs” for accelerators, networks, and scheduling tables, which serve as the primary input for system parameters.
Relationships & Data Flow
• Both branches should feed into “Common” components – that is, shared data structures and utility functions responsible for parsing, encapsulating design parameters, and error-checking.
Architectural Patterns and Design Principles
• Separation of concerns by clearly segmenting modules: the configuration/parsing layer, the mode-specific algorithms, and the shared utility components.
• That the design is modular, allowing the framework to support various configurations of accelerators and networks and plug in alternative mapping strategies.