Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

TVM v0.3 Release Note #854

Closed
tqchen opened this issue Jan 31, 2018 · 9 comments
Closed

TVM v0.3 Release Note #854

tqchen opened this issue Jan 31, 2018 · 9 comments
Labels
type: roadmap

Comments

@tqchen
Copy link
Member

tqchen commented Jan 31, 2018 •
edited
Loading

v0.3 is now tagged, next cycle roadmap issue is available at #1170

Release Note

This release features numerous improvements in TOPI and backends. We make the first step toward object detection support in TOPI, featuring operators necessary for YOLO and SSDs. The topi now supports numpy-style API and operator overloading. RPC is significantly improved to support resource allocation and using a pool of devices. We are adding two new backends: WebGL for running GPUs on the browser, and Vulkan for running on next-generation graphics API.

Change List

  • TOPI Vision operators
    • SSD support
    • YOLO support
    • NMS operator support in vision
  • TOPI general numpy-style operators
    • numpy style operator overload in topi
    • more operators: flip, take
    • dilation support on conv2d and depthwise
  • 8bit support
    • ARM 8bit gemm
    • ARM 8bit conv
  • Low bit operator support
    • popcount intrinsics
    • 1-bit fully connected
  • Contrib: MPSDNN fully-connected and conv2d support
  • Better RPC support
    • RPC Tracker support to allow centralized resource management
    • RPC protocol upgrade (this is a non-backward compatible change) to support timeout in the proxy
      • This is a breaking change, need to use the latest version of TVM runtime with the RPC
    • Fault-tolerant to early server termination with correct exception propagated
    • RPC support enabled for ROCm AMDGPUs
  • Tutorials and docs
    • How to deploy to android devices.
  • Optimizations for hardware backends
    • intel CPU (AVX and AVX512)
  • Schedule Primitives
    • rfactor now support factor_axis to specify the factored dimension in the result
    • cache_write now support multiple output operators
    • enable warp memory which generates shuffle instructions
  • Framework bridge
    • MXNet bridge supported
  • C++ compiler API support
    • build migration
    • topi migration to c++
    • Target system in c++
  • WebGL backend
    • runtime and codegen
    • topi integration
    • end to end pipeline on the browser
  • Vulkan backend
    • vulkan runtime
    • spirv code generator
  • Security
    • intel SGX runtime support
    • multi-threaded SGX runtime
  • LLVM 7.0 support
  • Robustness
    • VerifyMemory to verify incorrect GPU schedules that writes into GPU memory from cpu
    • Verify compute formulas
  • Better CPU parallel runtime

Main Contributors

See complete list here. Thanks to all the contributors to contribute to this release.

Code Reviewers

TOPI:

Compiler:
@tqchen tqchen mentioned this issue Jan 31, 2018
Closed
34 tasks
@yangjunpro
Copy link

Tianqi,

1-bit looks a cool feature. However, in our internal experiments, 1-bit weight(with 4-bit activation) still has non-negligible accuracy degradation(the best result still has around 3 percentage gap). So may I ask why we want to invest resource on this low-bit feature? Internally we are also leveraging TVM for boosting our productivity and already submit small patch to the community. We would like to align our engineering efforts with the community. Model compression is right now on our radars, and we would like to leverage TVM for fully exploiting the potential of model compressions.

Thanks.

@tqchen
Copy link
Member Author

tqchen commented Feb 1, 2018

We want to keep doors open for different kinds of optimizations, low bits(not necessarily 1bit) is among one of them. We are also thinking about more general low-bits(e.g. 2, 3 or 4 bits) operations. they allow interesting tradeoffs to be made(not necessarily).This direction can also be of particular interest to folks who works on hardware optimizations.

@starimpact
Copy link

8bit inference is the most important thing 😄

@danilopau
Copy link

Support for tensor comprehension would be great: optimized c code generation with openmp pragmas. Support for ONNX interoperability
Support for 32 FP down to 16FX, 8FX, 2 (ternary) and 1 bit (binary) weights and activations to enable vector instructions.

@tqchen
Copy link
Member Author

tqchen commented Apr 23, 2018

@danilopau Just so that you know that many things you mentioned are already supported. Include high performance optimized cpu code generation, ONNX interoperability, 32 FP to 16 FP, part of binary computation

@danilopau
Copy link

custom hardware backend example .... please look https://blog.st.com/orlando-neural-network-iot/.

@tqchen
Copy link
Member Author

tqchen commented May 14, 2018 •
edited
Loading

Thanks for everyone who have pushed to last release cycle in the past three month. I would like to propose release of v0.3 on May 21th. As usual, the current checklist will go into the release note, and we will move the unfinished ones into the roadmap of next release cycle.

Main contributors to the past release, see complete list here :
Thanks to all the contributors who makes the awesome progress

Code Reviewers

TOPI:

Compiler:

@tqchen
Copy link
Member Author

tqchen commented May 14, 2018

Please reply to see if there is something we would like to merge in before tag v0.3

@tqchen tqchen changed the title TVM v0.3 RoadMap TVM v0.3 Release Note May 21, 2018
@apache apache locked as resolved and limited conversation to collaborators May 21, 2018
@tqchen
Copy link
Member Author

tqchen commented May 21, 2018 •
edited
Loading

v0.3 is now tagged, next cycle roadmap issue is available at #1170

@tqchen tqchen closed this as completed May 21, 2018
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Assignees
No one assigned
Labels
type: roadmap
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
4 participants
@yangjunpro @starimpact @tqchen @danilopau