HLS Latency Predictor / HLS 延迟预测器

English · 中文

Production reproduction repository for pretrained-code HLS latency prediction and synthesis-free design-space exploration. The maintained implementation is under predictor/; data/ is read-only input and model weights are not stored in Git.

Validated target

• Known-workload DSE: validation-locked 12-expert ensemble, log_mae=0.1826, raw MAE about 252k cycles, pairwise accuracy 0.8913, NDCG@3 0.9771.
• Strict unseen-workload transfer: CodeT5-LoRA checkpoint, validation log_mae=0.6330 and test log_mae=1.9629.
• CHiRM/GNN-DSE/HARP-style synthesis-free loop: 75,000 2mm candidates in 145.38 s on the tested GPU; no HLS synthesis is invoked.

The DSE numbers are known-workload interpolation results and must not be used as strict unseen-workload claims.

Fast reproduction

git clone git@github.com:2021270902001sc/HLS_demo_code.git
cd HLS_demo_code
python -m pip install -r requirements.txt
./download_code.sh
python prepare_data.py --data data --output artifacts/manifest.jsonl

Train the production DSE components:

python train.py --config configs/qwen3_embedding_dse_signals.json \
  --data data --output artifacts/qwen3_embedding_dse_signals \
  --epochs 6 --batch-size 16 --device cuda:0

python scripts/train_static_expert.py --data data \
  --output artifacts/static_mlp_dse --strategy within_workload --device cuda:0
python scripts/train_static_expert.py --data data \
  --output artifacts/group_residual_mlp_dse --strategy within_workload \
  --split-seed 20260805 --seed 62 --group-feature group_id \
  --group-target residual --hidden-dim 512 --learning-rate 0.001 \
  --loss smooth --smooth-beta 0.1 --device cuda:0
python scripts/train_knn_expert.py --data data \
  --output artifacts/knn_static_dse --split-seed 20260805 --metric l1 --k 1
python scripts/train_meta_expert.py --data data \
  --qwen-train-predictions artifacts/qwen3_embedding_dse_signals/train_predictions.json \
  --qwen-validation-predictions artifacts/qwen3_embedding_dse_signals/validation_predictions.json \
  --qwen-test-predictions artifacts/qwen3_embedding_dse_signals/test_predictions.json \
  --qwen-checkpoint artifacts/qwen3_embedding_dse_signals/checkpoints/best \
  --output artifacts/meta_static_dse --device cuda:0

Run the complete predictor-only DSE loop:

python scripts/run_chirm_dse.py --mode Exhaustive \
  --point-limit 75000 --batch-size 64 --group-id 2mm:v20 \
  --device cuda:0 \
  --model static=artifacts/static_mlp_dse/checkpoints/best \
  --model residual=artifacts/group_residual_mlp_dse/checkpoints/best \
  --model knn=artifacts/knn_static_dse/checkpoints/best

--offline-data data/designs/v20/2mm.json is optional and is used only for post-hoc overlap checks; labels are never model inputs. The loop uses the local design-space definition at configs/chirm/2mm_ds_config.json and never calls transform, HLS, bitgen, or another synthesis command.

Repository contract

• predictor/: maintained data, model, training, evaluation, calibration and DSE implementation.
• configs/: production training and DSE configurations.
• scripts/run_chirm_dse.py: complete dependency-aware, batch, cached DSE.
• data/ and demo_data/: supplied input datasets; never rewritten by tools.
• tests/: regression tests (python -m pytest -q).
• models/, encoder_codet5/, artifacts/ and reports/: local-only files; weights and generated runs are intentionally excluded from Git.

For exact historical research notes and intermediate experiments, use the separate private repository HLS_demo_code_private.