TVM v0.2 Release Note

[tqchen]

Thanks for everyone for a good first release, this issue contains the road-map and checklist we want to achieve in 0.2, feel free to comment, we will open separate issues when necessary for detailed progress checking. Feel free to volunteer yourself if you are interested in trying out some items.

As usual, the checklist will go into the release note.

Features

This release comes with a complete set of TOPI support for NNVM compiler, which allows compilation of end to end workloads. We also make major improvements in supporting new backends: ROCm for AMDGPUs and optimized ARM GPU performance.

• Backend support
  • Support LLVM mainline(4.0, 5.0, 6.0)
  • Support ROCM stack for AMD GPUs
  • WebGL backend
    • Basic runtime and compiler
    • topi
    • end to end example
• C++ compiler support
  • Basic version of build function in C++
  • migration of functions
• Runtime
  • Android RPC @Javelinjs
• Multi-threading optimization
  • multi-threaded depthwise
  • multi-threaded conv2d
• Schedule primitives
  • storage_align
  • double_buffer
• End to end pipelines
  • CPU ResNet18
  • CPU MobileNet
  • GPU ResNet
• TOPI operators
  • broadcast/ reduction operators
  • pooling and global pooling
  • Generic target support for topi
  • schedule with external libraries
• Target system
  • Introduce tvm.target to better specify target options for compilation.
• Tutorials
  • How to load compiled module in any language runtime
  • How to use java runtime
• Contrib library
  • MIOpen
  • MPS DNN
    • Dense
    • convolution
• 8bit support
  • ARM 8bit gemm @ZihengJiang
  • ARM 8bit conv @ZihengJiang
  • Pascal 8bit gemm
• One bit support
  • popcount intrinsics
  • 1bit convolution @Huyuwei
• ARM GPU support
  • gemm
  • conv2d
  • depthwise
  • end to end pipeline

[Huyuwei]

Great interest in 1bit convolution!

[Laurawly]

Gonna finish the GPU end to end Resnet and interested in gpu optimization on Volta.

[ZihengJiang]

I would like to take 8bit on ARM

[yzhliu]

I'd like to take gpu on android device.

[tqchen]

also cc @ysh329 who is also interested in android gpu opencl devices.

[csarron]

May also consider RenderScript implementation for GPU on Android. @Javelinjs

[ajtulloch]

For low bitwidth conv/gemm implementations, you might find https://github.com/caffe2/caffe2/blob/master/caffe2/mobile/contrib/ulp2/ulp_neon.cc interesting as a baseline. The gess microkernel hits >80% of peak on a few common microarchitectures on ARMv7.

[ajtulloch]

And re: wrapping MPS in C++, https://github.com/caffe2/caffe2/blob/master/caffe2/mobile/contrib/ios/mpscnn/mpscnn.mm might be interesting as well.

[tqchen]

thanks @ajtulloch the links are helpful. It might make sense to put the neon micro kernel as contrib so we can use RPC to test it in a integrated fashion

[tqchen]

It has been five months a year since v0.1 release and we have made a lot of progress. I would like to propose release of v0.2 on Jan 31th. The current checklist will go into the release note, and we will move the unfinished ones into the roadmap of next release cycle.

Thanks everyone's effort to put in the great set of improvements.

[antinucleon]

Feb 31?

On Wed, Jan 24, 2018 at 12:19 Tianqi Chen ***@***.***> wrote: It has been five months a year since v0.1 release and we have made a lot of progress. I would like to propose release of v0.2 on Feb 31th. The current checklist will go into the release note, and we will move the unfinished ones into the roadmap of next release cycle. Thanks everyone's effort to put in the great set of improvements. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#349 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABM13snrP1zcgKx_BlK6SIsS0Naic_EJks5tN5BXgaJpZM4O8NUC> .

-- Sent from mobile phone

[tqchen]

Sorry should be Jan 31th, corrected in the comment

[tqchen]

#853

[tqchen]

v0.2 has been officially tagged. close this issue. The next cycle roadmap is at #854