MXNetOnACL is a project that is maintained by OPEN AI LAB, it uses Arm Compute Library (NEON+GPU) to speed up MXNet and provide utilities to debug, profile and tune application performance.
The release version is 0.2.0, is based on Rockchip RK3399 Platform, target OS is Ubuntu 16.04. Can download the source code from OAID/MXNetOnACL
- The ARM Computer Vision and Machine Learning library is a set of functions optimised for both ARM CPUs and GPUs using SIMD technologies. See also Arm Compute Library.
- MXNet is a Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more. See also MXNet.
There are some compatibility issues between ACL and Caffe Layers, we bypass it to Caffe's original layer class as the workaround solution for the below issues
- Normalization in-channel issue
- Tanh issue
- Softmax supporting multi-dimension issue
- Group issue
Performance need be fine turned in the future
The MXNet based version is 26b1cb9ad0bcde9206863a6f847455ff3ec3c266.
Support Arm Compute Library version 17.06 with 4 new layers added
- Batch Normalization Layer
- Direct convolution Layer
- Concatenate layer
Initial version supports 10 Layers accelerated by Arm Compute Library version 17.05 :
- Convolution Layer
- Pooling Layer
- LRN Layer
- ReLU Layer
- Sigmoid Layer
- Softmax Layer
- TanH Layer
- AbsVal Layer
- BNLL Layer
- InnerProduct Layer
Encounter any issue, please report on issue report. Issue report should contain the following information :
- The exact description of the steps that are needed to reproduce the issue
- The exact description of what happens and what you think is wrong