IntrinTrans is a LangGraph-based agentic pipeline that automatically migrates ARM NEON vector intrinsic C/C++ code to RISC-V Vector (RVV) intrinsic code. It combines an LLM translation agent with iterative compile-test-optimize feedback loops to produce correct, vectorized RVV implementations.
The pipeline is implemented as a LangGraph StateGraph with the following nodes:
START
│
▼
translator ──────────────────────────────────────────────┐
│ │ (compile/test failure → fix)
▼ │
compiler ──(fail)──► translator (fix) / select_best ◄────┘
│
▼ (success)
tester ──(fail)──► translator (fix) / optimizer (fix) / select_best
│
▼ (pass)
optimizer ──► compiler ──► tester
│ │
│ (pass)▼
│ record_version
│ │
└─────────────── (loop) ────┘
│ (limit reached)
▼
select_best ──► END
| Node | Responsibility |
|---|---|
| translator | Calls LLM to generate or fix RVV code from the ARM NEON source; manages multi-turn conversation history |
| compiler | Cross-compiles the generated code with CMake + Ninja; runs optional register liveness analysis |
| tester | Executes the binary via QEMU or remote SSH; validates [Tests Passed] and checks for RVV intrinsic usage |
| optimizer | Calls LLM to improve register usage guided by liveness analysis; supports fix sub-loops |
| record_version | Appends each passing optimized variant to all_passed_versions |
| select_best | Picks the highest-speedup (or latest) version as best_version |
| Requirement | Notes |
|---|---|
| Python ≥ 3.9 | Async/await, TypedDict generics |
| RISC-V GCC toolchain | Cross-compiler targeting rv64gcv; set RISCV_TOOLCHAIN_PATH |
| CMake ≥ 3.16 + Ninja | Build system for cross-compilation |
QEMU (qemu-riscv64) |
For local functional testing at vlen=128/256; or a remote RISC-V device |
| OpenAI-compatible API key | Any provider routed via OPENAI_BASE_URL (e.g., OpenRouter) |
| VecIntrinBench | Benchmark suite providing ARM NEON source files and GTest-based test files (included as a git submodule) |
# 1. Clone the repository git clone git@github.com:RuyiAI-Stack/IntrinTrans.git cd IntrinTrans # 2. Initialize submodules git submodule update --init --recursive # 3. Install Python dependencies pip install -r requirements.txt # 4. Configure environment cp .env.example .env # Edit .env with your values # 5. Validate configuration python run.py --check-config
python run.py --func-name <casename> --model <model-id>
The tool auto-locates source and test files from:
VecIntrinBench/src/arm/<casename>.cpp— ARM NEON sourceVecIntrinBench/test/<casename>.cpp— GTest-based functional testVecIntrinBench/perf/<casename>.cpp— (optional) performance test
Examples:
# Translate using DeepSeek V3 python run.py --func-name volk_32u_popcnt --model deepseek/deepseek-v3.2 # Translate with custom attempt limits python run.py --func-name volk_32u_popcnt \ --model google/gemini-3.1-pro-preview \ --max-translation 5 \ --max-optimization 2
VITrans-CSCWD/
├── run.py # CLI entry point
├── requirements.txt # Python dependencies
├── .env.example # Environment configuration template
│
├── intrintrans/ # Core package
│ ├── config.py # Config loader (.env + os.environ)
│ ├── state.py # LangGraph state types (TranslationState, TranslationPhase)
│ ├── graph.py # LangGraph workflow definitions (full / opt-only)
│ │
│ ├── agents/
│ │ └── llm_client.py # Async OpenAI-compatible client with retry logic
│ │
│ ├── nodes/
│ │ ├── translator.py # Translator node (NEON → RVV via LLM)
│ │ ├── compiler.py # Compiler node (CMake + Ninja cross-compilation)
│ │ ├── tester.py # Tester node (QEMU / remote SSH execution)
│ │ └── optimizer.py # Optimizer, record_version, select_best nodes
│ │
│ └── utils/
│ ├── build_tools.py # cmake_build_and_test, run_test, parse_code, liveness analysis
│ └── logging_utils.py # Step record helpers, log formatting
│
└── VecIntrinBench/ # Benchmark submodule (ARM NEON sources + GTests)
├── src/arm/ # Input: ARM NEON intrinsic functions
├── test/ # Functional test cases (GTest)
├── perf/ # Performance test cases (optional)
└── toolchain/riscv.cmake # RISC-V CMake toolchain file