Board Environment Dependencies

This demo is intended to run on the AidLux board environment where the AidLite QNN240 runtime is available. A normal development container can run --dry_run, but cannot execute the QNN contexts unless import aidlite succeeds.

Verified Board Environment

The current board used for validation reports:

Python 3.10.12
aidlite import: OK
numpy 1.26.4
opencv-python / cv2 4.13.0
Pillow / PIL 10.4.0

Required Runtime Components

AidLux / AidLite Python runtime with aidlite module.
AidLite QNN240 plugin/runtime for FrameworkType.TYPE_QNN240.
Qualcomm HTP/DSP runtime and valid board license.
Board-side QNN context execution support for the selected model platform.

The Python code checks the AidLite enum contract at startup:

FrameworkType.TYPE_QNN240 = 109
ImplementType.TYPE_LOCAL = 3
AccelerateType.TYPE_DSP = 3

Required Python Packages

The demo imports these Python packages at runtime:

numpy
cv2
PIL
aidlite

Package usage:

aidlite: loads and invokes the four QNN240 context .aidem files.
numpy: tensor loading, dtype conversion, BEVFormer postprocess, NPZ output.
cv2: board-side six-camera JPG preprocessing.
PIL: camera-grid PNG/GIF visualization.

Model/Data Files Required By Default Run

Default command:

python3 code/python/run_test.py --invoke_nums 4 --output_dir outputs/final_sample4

The default command auto-discovers the first complete model directory in this order:

models/QCS8550/FP16
models/IQ9/FP16
models/QCS8625/FP16
models/IQ8/FP16

Model SHA verification is always enabled. The platform is inferred from model paths:

QCS8550 -> QCS8550 model SHA table
IQ9     -> QCS8550 model SHA table
QCS8625 -> QCS8625 model SHA table
IQ8     -> QCS8625 model SHA table

Each platform directory must contain:

backbone_context.bin.aidem
scene_start_encoder_context.bin.aidem
temporal_encoder_context.bin.aidem
decoder_context.bin.aidem

Required config files:

code/python/configs/demo_config.json
code/python/configs/nms_runtime_contract.json

Required sample data:

code/python/datasets/sample4/asset_manifest.json
datasets/sample4/frames/sample_000..sample_003/

Each sample frame contains six raw camera JPGs plus the small auxiliary tensors required by the deployed contexts.

Quick Environment Check

Run on the board:

cd /home/aidlux/bevformer_delivery_demo
python3 -c "import aidlite, numpy, cv2; from PIL import Image; print('board env ok')"
python3 code/python/run_test.py --dry_run --check_raw_assets

The dry-run checks paths, model SHA, sample assets, and scene-start/temporal routing. It does not invoke AidLite/DSP.

Notes

Do not expect real inference to work in a host/container environment without AidLite.
configs/qnn_htp_configs/ is not required by runtime; it was only an optional encryption/audit supplement and is not included in the runtime demo package.
Full inference timing and visualization are printed by the default command without adding extra timing flags.