How set-scheduling-project/gemini-hpca2024 works
Project type and purpose
• This is a research-grade command-line tool (Gemini) for design-space exploration (DSE) of large-scale DNN chiplet accelerators.
• It is implemented in C++ (core mapping engine) with Python orchestration scripts for batch DSE and result post-processing.
• It reads front-end IRs (JSON descriptions of DNNs), explores mapping/partition/placement/spatial-mapping strategies, computes cost models, and outputs performance/cost results.
File structure
• Top-level:
• .sh scripts (e.g., 72tops_dse13.sh): orchestrate parameter sweeps.
• makefile: builds the C++ binary.
• src/, include/: core engine sources and headers.
• src/front_end_IRs/: sample DNN JSON models (e.g., SSD, VGG16).
• pyscripts/: Python utilities for converting logs to CSV/XLSX and for “best architecture” selection.
• Key modules under include/ & src/:
• IR parsing (json/, front_end_IRs)
• Network definitions (nns/)
• Core mapping (coremapping), clustering (cluster), data-layout (datalayout)
• Partitioning, placement, spatial mapping (spatial_mapping/DP, light_placement, segmentation)
• Cost modeling (cost, bufferusage, packaging cost via references)
• NoC modeling (noc)
• Main driver (main.cpp)
• Configuration/inputs: JSON IRs, .sh DSE scripts, Python requirements.txt.
README for additional insights
• Describes supported process nodes, DRAM types, D2D interfaces, packaging technologies.
• Explains build/run steps: make, pip install -r requirements.txt, run a DSE shell script.
• Notes that foundational code is shared with the SET framework.
a. Main components
1. DSE Orchestrator (Shell scripts + Python wrappers)
2. C++ Mapping Engine (CLI executable)
3. Front-end IR Loader (JSON parser + IR models)
4. Mapping Pipeline:
– Partition Module
– Placement Module
– Spatial-Mapping Module
– Core-Mapping & Clustering Module
5. Cost & Performance Model (cost, bufferusage, NoC)
6. Result Generator (log files)
7. Post-Processing Python Tools (csv2xlsx, log2csv, best_arch.py)
b. Relationships and interactions
• The DSE Orchestrator invokes the C++ engine repeatedly with varying parameters.
• The C++ engine reads a JSON IR, runs pipeline modules in sequence, writes textual logs.
• Python post-processors consume log files, produce CSV/XLSX summaries and best-arch selection.
c. Architectural patterns/design principles
• Pipeline pattern: sequential modules transforming an IR.
• Plugin-style front-end: new networks added via JSON and nns/ implementations.
• File-based orchestration: loose coupling between DSE driver, engine, and analysis tools.
d. Technologies/frameworks
• C++17, Makefile build
• JSON for Modern C++ (include/json, src/json)
• Python 3 for orchestration and post-processing
• Shell scripting for parameter sweeps