Mobile AI Compute Engine (or MACE for short) is a deep learning inference framework optimized for mobile heterogeneous computing platforms. The design focuses on the following targets:
- Runtime is optimized with NEON, OpenCL and Hexagon, and Winograd algorithm is introduced to speed up convolution operations. The initialization is also optimized to be faster.
- Power consumption
- Chip dependent power options like big.LITTLE scheduling, Adreno GPU hints are included as advanced APIs.
- UI responsiveness guarantee is sometimes obligatory when running a model. Mechanism like automatically breaking OpenCL kernel into small units is introduced to allow better preemption for the UI rendering task.
- Memory usage and library footprint
- Graph level memory allocation optimization and buffer reuse are supported. The core library tries to keep minimum external dependencies to keep the library footprint small.
- Model protection
- Model protection has been the highest priority since the beginning of the design. Various techniques are introduced like converting models to C++ code and literal obfuscations.
- Platform coverage
- Good coverage of recent Qualcomm, MediaTek, Pinecone and other ARM based chips. CPU runtime is also compatible with most POSIX systems and architectures with limited performance.
MACE Model Zoo contains several common neural networks and models which will be built daily against a list of mobile phones. The benchmark results can be found in the CI result page (choose the latest passed pipeline, click release step and you will see the benchmark results).
- GitHub issues: bug reports, usage issues, feature requests
- Slack: mace-users.slack.com
- QQ群: 756046893
Any kind of contribution is welcome. For bug reports, feature requests, please just open an issue without any hesitation. For code contributions, it's strongly suggested to open an issue for discussion first. For more details, please refer to the contribution guide.
MACE depends on several open source projects located in the third_party directory. Particularly, we learned a lot from the following projects during the development:
- Qualcomm Hexagon NN Offload Framework: the Hexagon DSP runtime depends on this library.
- TensorFlow, Caffe, SNPE, ARM ComputeLibrary, ncnn and many others: we learned many best practices from these projects.
Finally, we also thank the Qualcomm, Pinecone and MediaTek engineering teams for their help.