Blog of the TVM deep learning compiler project, covering GPU backends, operator optimization, and deployment on mobile and web targets.
HN commenters link it as a source on TVM's broad hardware support, including AMD GPUs, and its role next to ONNX and dlpack.
A collection about depthwise conv:
Optimize Deep Learning GPU Operators with TVM: A Depthwise Convolution Example
http://tvmlang.org/2017/08/22/Optimize-Deep-Learning-GPU-Operators-with-TVM-A-Depthwise-Convolution-Example.html
MobileNet and Depthwise Separable Convolution · Issue #70 · changtimwu/changtimwu.github.com
https://github.com/changtimwu/changtimwu.github.com/issues/70
yonghenglh6/DepthwiseConvolution: A personal mobile convolution layer implementation on caffe by liuhao.(only GPU)
https://github.com/yonghenglh6/DepthwiseConvolution
Julia is a good language for computational simulation experiments but in a world that's heading towards heterogeneous compute enabled by GPUs, TPUs and intermediate representation languages/frameworks like ONNX, TVM [1] and Tensor Flow its difficult to take claim of "Language for Modern ML" seriously.
If you disagree please show me how many ML researchers/labs/companies use Julia over Python/C++. Its cool to claim "Modern", "Deep", "ML", but I don't see any evidence.
[1] http://tvmlang.org/2017/08/17/tvm-release-announcement.html
It looks like the same group that is behind MXNet recently released TVM [1], which lets you use Cuda, OpenCL, Metal, etc as the target for tensor operations. I wonder if that means support for AMD cards will happen in the near future?
1. http://www.tvmlang.org/2017/08/17/tvm-release-announcement.html
Recently there is more and more initiative to have standard format in deeplearning environnement.
dlpack for tensor format (https://github.com/dmlc/dlpack)
onnx for saved NN (https://github.com/onnx/onnx)
and tvm for execution (http://tvmlang.org/2017/10/06/nnvm-compiler-announcement.html)
Shameless plug for the tvm project: http://tvmlang.org/2017/10/30/Bringing-AMDGPUs-to-TVM-Stack-and-NNVM-Compiler-with-ROCm.html which is an effort towards broad (many vendors) hardware support and optimization for deep learning.