- New work flow and documents.
- Separate the model library from MACE library.
- Reduce the size of static and dynamic library.
- Support
ArgMax
Operations. - Support
Deconvolution
of Caffe. - Support NDK-17b.
- Use file to store OpenCL tuned parameters and Add
SetOpenCLParameterPath
API.
- Add a new
MaceEngine::Init
API with model data file.
- Not unmap the model data file when load model from files with CPU runtime.
- 2D LWS tuning does not work.
- Winograd convolution of GPU failed when open tuning.
- Incorrect dynamic library of host.
Appreciate for the following guys contribute code to make MACE better.
Zero King(@l2dy), James Bie(@JamesBie), Sun Aries(@SunAriesCN), Allen(@allen0125), conansherry(@conansherry), 黎明灰烬(@jackwish)
- Change build and run tools
- Handle runtime failure
- Change interface that report error type
- Improve CPU performance
- Merge CPU/GPU engine to on
- support
float
data_type
when running in GPU
- Return status instead of abort when allocate failed
- Change mace header interfaces, only including necessary methods.