TVM Monthly - April 2021

刺骨的言语ヽ痛彻心扉 2022-09-09 14:59 163阅读 0赞

# [TVM Monthly - April 2021][] #

As discussed by the TVM PMC, our goal is to provide a monthly summary of the project so users and developers can get a better understanding of the goings on of the TVM community.

Feedback and suggestions are welcomed so that we can further improve these updates.

## Community ##

During April of 2021 we welcomed many new contributors to the project. Importantly we welcomed [@kevinthesun][kevinthesun] as a PMC member, [@Meteorix][Meteorix], [@Hzfengsy][Hzfengsy], [@spectrometerHBH][spectrometerHBH], and [@giuseros][giuseros] as reviewers.

Thanks to everyone for their hardwork and contributions!

On the technical side we kept working on improving operator and frontend support, especially a large set of ONNX support has been landed. As have been discussed in [TensorIR: A schedulable IR for TVM][TensorIR_ A schedulable IR for TVM], we are incrementally adding features to this new type of tensor-level IR in TVM. The entire progress is tracked in [TensorIR Scheduling Tracking Issue 3][TensorIR Scheduling Tracking Issue_3]. MicroTVM has been enhanced and supports Zephyr on RISC-V QEMU. Moreover, we further improved the passes in Relay and provided a unified interface for profiling between the graph runtime and Relay VM. The community has also improved documents and tutorials a lot.

This forum got 122k pageviews, 3.0k user visits in the last month.

## Relay ##

*  \[PatternMatcher\] Support matching tuples, call nodes, and functions with variable numbers of inputs [\#7754][7754]
 *  RelayTextPrinter is now non-recursive. ExpandDataflow refactored [\#7817 1][7817_1]
 *  \[Relay\] Recursive destructor call replaced with non-recursive for Call nodes. [\#7832][7832]
 *  \[Relay\]\[Pass\] Avoid stack overflow when using PostOrderRewrite [\#7588][7588]
 *  \[Relay\]\[Pass\] SimplifyCastLike/Cast and ConcretizeFullLikeRewrite rewrites for SimplifyExpr [\#7827 1][7827_1]
 *  \[Relay\] Add support for relay expressions as pad value for static pad [\#7860][7860]
 *  \[ConvertLayout\] Keep span in ConvertLayout [\#7895][7895]
 *  \[ConvertLayout\] Squeeze and reduce ops [\#7835][7835]
 *  \[Topi & Relay\] Add quantization support for the vision transform model in GPU [\#7814 1][7814_1]
 *  \[Relay\]\[Op\] Logical Not Shape Function [\#7820 1][7820_1]

## Tensor-level IR and Arithmetics ##

*  \[TIR\] Add a new intrinsic count leading zeros for LLVM and SPIR-V [\#7825 2][7825_2]
 *  \[TIR\]\[TRANSFORM\] Return value support in tir.tvm\_call\_packed [\#7932][7932]
 *  \[TensorIR\]\[PASS\]\[M1c\] PlanUpdateBufferAllocationLocation [\#7873][7873]
 *  \[TensorIR\]\[M1c\] LCA detector [\#7848][7848]
 *  \[TensorIR\]\[M1c\] LowerInitBlock [\#7806][7806]
 *  \[TensorIR\]\[PASS\]\[M1c\] CompactBufferAllocation [\#7923][7923]
 *  \[TensorIR\] \[Script\] adding support for opaque block [\#7829][7829]
 *  \[TensorIR\]\[M1b\] Scaffolding ScheduleState data structure [\#7765][7765]
 *  \[TensorIR\]\[M1b\] Schedule class [\#7847 1][7847_1]
 *  \[TIR\] An analysis pass to calculate workspace size for primfuncs [\#7859][7859]
 *  \[ARITH\] Subspace division [\#7760][7760]

## Language Bindings ##

*  \[Rust\] Make TVM Rust bindings installable via Cargo. [\#7503][7503]

## Frontend ##

### ONNX ###

*  \[ONNX\] Support optional outputs for ONNX nodes [\#7818][7818]
 *  \[ONNX\] Initial work to import pre-quantized ONNX Models [\#7802 1][7802_1]
 *  \[ONNX\] Dynamic Gather [\#7787][7787]
 *  \[ONNX\]\[TOPI\]\[RELAY\] Resize refactor [\#7883][7883]
 *  \[ONNX\] Fix more upstream tests [\#7842][7842]
 *  \[ONNX\] Support importing Conv with missing attributes [\#7899][7899]
 *  \[ONNX\] Support NMS Center Box [\#7900][7900]
 *  \[ONNX\] Bitshift Operator [\#7800][7800]
 *  \[TOPI\]\[RELAY\]\[ONNX\] Scatter ND [\#7927 1][7927_1]
 *  \[TOPI, Relay\] A new NMS op variant for ONNX NMS / TF Combined NMS [\#7796 1][7796_1]
 *  \[Relay\]\[ONNX\] 1-D global and adaptive pooling. [\#7906][7906]
 *  \[Relay\]Frontend\]\[Onnx\] Remove pop that interferes with nested loops. [\#7781][7781]
 *  \[ONNX\] Collect quant params of pre-quantized ONNX and generate qnn op [\#7937 1][7937_1]
 *  \[ONNX\] Make input shape immutable [\#7844][7844]

### Tensorflow and TFLite ###

*  \[Frontend\]\[Tensorflow\] Support SAME padding for dynamic h, w when stride == 1 [\#7885][7885]
 *  \[QNN\]\[TFLite\] TFLite rounding mode support [\#7456 3][7456_3]
 *  \[frontend\]\[tflite\] float16 quant support [\#7736 2][7736_2]
 *  Quantized RESIZE\_BILINEAR operator support in TF Lite Frontend [\#7866][7866]

## Backend ##

*  \[TOPI\]\[SPIRV\] Cast to float32 not float64 before log2 in sort/scan [\#7669][7669]
 *  \[Vulkan\] Support uniform buffer object for passing many scalar arguments (Take 2) [\#7833][7833]
 *  \[Hexagon\] Reenable compilation of TVM runtime for Hexagon [\#7784 1][7784_1]
 *  \[CodeGen\]\[OpenCL\] Limit OpenCL built-in vector lanes to 2, 3, 4, 8, 16. [\#7777][7777]

## Compilation Flow and API ##

*  \[TVMC\] --disable-pass option added to compile mode [\#7816][7816]
 *  \[Target\] Add support for target object with host field compatible with previous api [\#7534][7534]
 *  \[Target\]\[Lowering\] Update Op Intrinsic Lowering Mechanism And Intrinsic Lowering Pass [\#7809][7809]
 *  \[TVMC\] Separate model loading from model compilation in TVMC. [\#7739][7739]
 *  \[TVMC\] Allow direct numpy inputs to run\_module [\#7788][7788]
 *  \[TVMC\] Autotuning - Hardware configs not default [\#7792][7792]

## Ansor, Autoscheduler and AutoTVM ##

*  Protect child process enumeration in AutoTVM [\#7887][7887]
 *  Enable StackVM in AutoTVM [\#7897][7897]
 *  \[AutoScheduler\] Add task.desc for its function name [\#7794 1][7794_1]

## MicroTVM ##

*  \[microTVM\] Add error reporting module to microTVM runtime [\#7671][7671]
 *  \[microTVM\] Update nrfjprog on reference virtual machine [\#7723 ][7723]
 *  \[microTVM\] Zephyr: RISCV support for Zephyr QEMU RISCV-32/64 [\#7804 2][7804_2]
 *  \[µTVM\] Zephyr: Add STM32F746 disco board as a test platform [\#7863 ][7863]
 *  \[µTVM\] Clone Zephyr 2.5.0 from maintenance branch [\#7891][7891]
 *  \[µTVM\] Add support for mps2\_an521 board [\#7813][7813]
 *  \[µTMV\] apps: Fix Zephyr code example for STM32F746 boards [\#7772][7772]
 *  \[µTVM\] Zephyr: Add MPS2-AN521 board as a test platform [\#7864][7864]
 *  Allow microTVM Reference VM to be launched when TVM is a submodule. [\#7854][7854]

## Runtime ##

*  \[Profiling,VM\] Profiling interface for VM and Graph Runtime [\#7624 2][7624_2]
 *  \[PROFILER\] Add CSV output to profiler [\#7797 3][7797_3]
 *  \[Runtime\] Driver version + consistent clock speed units [\#7867][7867]
 *  \[Refactor\]\[VM\] Port memory\_alloc to c++ [\#7369][7369]
 *  \[RUNTIME\] Add clear() function in tvm::Map class [\#7826][7826]

## Build and CI ##

*  \[CI\] v0.7 CI updates [\#7947][7947]
 *  Rev ci-qemu container to v0.04 [\#7946][7946]
 *  \[µTVM\] Try to fix qemu hangs in the CI \#7590 [\#7769][7769]
 *  \[CI\] Rust CI Changes [\#7773][7773]
 *  Disable Rust CI [\#7793][7793]
 *  \[CI\] add the --net=host cmd line arg to the docker/bash.it script [\#7780][7780]
 *  \[CI\] docker images build script cmd line args optional [\#7776][7776]
 *  Use new SBT Debian Repo before bintray is shutdown [\#7926][7926]

## Doc ##

*  \[docs\] Getting Started With TVM: Tensor Expressions [\#7768 ][7768]
 *  \[docs\] Getting Started with TVM: Auto Tuning with Python [\#7767][7767]
 *  \[DOCS\] Remove stale Auto TensorCore CodeGen tutorial [\#7924 ][7924]
 *  \[DOCS\]\[BUILD\] Small improvements to documentation/build setup for first-time builds [\#7840][7840]
 *  \[Docs\] Update Dev. Doc. on how to add a new relay operator [\#7893][7893]
 *  \[RELEASE\] Update NEWS.md for v0.7 [\#6613][6613]
 *  Update ICHECK error message with link to documentation page. [\#7869][7869]
 *  \[DOC\] Grammar fix [\#7824][7824]

## Improvement and Bugfix ##

*  \[TIR\]\[SPIR-V\] Fix computing clz on int64 input for vulkan [\#7913][7913]
 *  \[Relay\] Shape func fix for all\_class\_nms and where op [\#7910][7910]
 *  \[Hotfix\] Typo in Vulkan runtime change causing severe perf regression for dGPU [\#7871][7871]
 *  \[RUNTIME\]\[BUGFIX\] Fix DSO module problem when its parent get destructed. [\#7918][7918]
 *  \[RPC\]\[REFACTOR\] Use PopenWorker to handle RPC Server. [\#7889][7889]
 *  \[VTA\] Update vta-hw dependency [\#7874][7874]
 *  fix compiling warning in simplify\_expr.h [\#7828][7828]
 *  \[RPC\] microtvm: fix RPC large transfer size issue [\#7838][7838]
 *  \[microTVM\] Update Zephyr 2.5 [\#7786 1][7786_1]
 *  \[microTVM\] Refactor zephyr installation + Update Zephyr RVM doc [\#7915][7915]
 *  \[microTVM\] Fix RVM onnx dependency and Zephyr document update [\#7774][7774]
 *  \[FIX\] Fix howto\_deploy [\#7841][7841]
 *  \[FIX\] tvm.testing.parametrize\_targets documentation for arguments does not match what it is acutally using [\#7778][7778]
 *  \[FIX\] Make HashCombine stable across platforms [\#7801][7801]
 *  \[FIX\] Fix RPC for the VM [\#7810][7810]
 *  \[Tophub\] Race condition fixed in folder creation [\#7940][7940]
 *  \[BYOC\]\[TVMC\] bugfix: disabled\_pass → disable\_pass [\#7850][7850]
 *  \[BYOC\]\[ACL\] ACL migrated to v21.02 [\#7649][7649]
 *  Fix typos in comments [\#7862][7862]
 *  Bring back GraphRuntimeFactory loader for now. [\#7868][7868]
 *  Fix Zephyr flashing on physical hardware, busted in \#7813 [\#7853][7853]
 *  \[Target\] Update tags with minor fix [\#7448][7448]
 *  \[Target\] Fix empty target and host for autotvm task [\#7791][7791]
 *  \[TE\] Fix bug if find a loop in compute\_at attach path [\#7898][7898]
 *  \[TE\] Bugfix for reduction that involves multi-outs with where cond [\#7692][7692]
 *  \[CodeGenC\] Fix bugs when calling extern functions [\#7911][7911]
 *  \[Tensorize\] Fix compute reusing [\#7920][7920]
 *  \[TVMC\] Runner.py Updates [\#7779][7779]
 *  \[autodiff\] change `truncdiv` to `div` in EliminateDivModMutator [\#7943][7943]
 *  Remove pointer arithmetic in StorageObj::AllocNDArray [\#7890][7890]
 *  Fix compilation errors with clang 11 [\#7783][7783]
 *  \[ROCM\]\[Bugfix\] Thread local handle for rocblas [\#7851][7851]
 *  \[TF frontend\]\[bugfix\]Avoid making a new node when already has span info [\#7789][7789]
 *  \[BugFix\] Print doubles with precision 17 in SaveJSON and TVM script printer [\#7846][7846]
 *  Remove unnecessary bracket around make\_int [\#7907][7907]
 *  Fix PyTorch matmul conversion when given (2-dim, N-dim) input pair [\#7845 ][7845]
 *  Added missing include file [\#7808][7808]
 *  Fix Metal accuracy problem caused by 3 vectors usage [\#7830 ][7830]
 *  \[BugFix\]: Convert tuple to int [\#7880][7880]
 *  \[FIX\] skip\_conv\_layers will affect quantization of nn.dense [\#7795][7795]

[TVM Monthly - April 2021]: https://discuss.tvm.apache.org/t/tvm-monthly-april-2021/9937
[kevinthesun]: https://discuss.tvm.apache.org/u/kevinthesun
[Meteorix]: https://discuss.tvm.apache.org/u/meteorix
[Hzfengsy]: https://discuss.tvm.apache.org/u/hzfengsy
[spectrometerHBH]: https://discuss.tvm.apache.org/u/spectrometerhbh
[giuseros]: https://discuss.tvm.apache.org/u/giuseros
[TensorIR_ A schedulable IR for TVM]: https://discuss.tvm.apache.org/t/rfc-tensorir-a-schedulable-ir-for-tvm/
[TensorIR Scheduling Tracking Issue_3]: https://github.com/apache/tvm/issues/7527
[7754]: https://github.com/apache/tvm/pull/7754
[7817_1]: https://github.com/apache/tvm/pull/7817
[7832]: https://github.com/apache/tvm/pull/7832
[7588]: https://github.com/apache/tvm/pull/7588
[7827_1]: https://github.com/apache/tvm/pull/7827
[7860]: https://github.com/apache/tvm/pull/7860
[7895]: https://github.com/apache/tvm/pull/7895
[7835]: https://github.com/apache/tvm/pull/7835
[7814_1]: https://github.com/apache/tvm/pull/7814
[7820_1]: https://github.com/apache/tvm/pull/7820
[7825_2]: https://github.com/apache/tvm/pull/7825
[7932]: https://github.com/apache/tvm/pull/7932
[7873]: https://github.com/apache/tvm/pull/7873
[7848]: https://github.com/apache/tvm/pull/7848
[7806]: https://github.com/apache/tvm/pull/7806
[7923]: https://github.com/apache/tvm/pull/7923
[7829]: https://github.com/apache/tvm/pull/7829
[7765]: https://github.com/apache/tvm/pull/7765
[7847_1]: https://github.com/apache/tvm/pull/7847
[7859]: https://github.com/apache/tvm/pull/7859
[7760]: https://github.com/apache/tvm/pull/7760
[7503]: https://github.com/apache/tvm/pull/7503
[7818]: https://github.com/apache/tvm/pull/7818
[7802_1]: https://github.com/apache/tvm/pull/7802
[7787]: https://github.com/apache/tvm/pull/7787
[7883]: https://github.com/apache/tvm/pull/7883
[7842]: https://github.com/apache/tvm/pull/7842
[7899]: https://github.com/apache/tvm/pull/7899
[7900]: https://github.com/apache/tvm/pull/7900
[7800]: https://github.com/apache/tvm/pull/7800
[7927_1]: https://github.com/apache/tvm/pull/7927
[7796_1]: https://github.com/apache/tvm/pull/7796
[7906]: https://github.com/apache/tvm/pull/7906
[7781]: https://github.com/apache/tvm/pull/7781
[7937_1]: https://github.com/apache/tvm/pull/7937
[7844]: https://github.com/apache/tvm/pull/7844
[7885]: https://github.com/apache/tvm/pull/7885
[7456_3]: https://github.com/apache/tvm/pull/7456
[7736_2]: https://github.com/apache/tvm/pull/7736
[7866]: https://github.com/apache/tvm/pull/7866
[7669]: https://github.com/apache/tvm/pull/7669
[7833]: https://github.com/apache/tvm/pull/7833
[7784_1]: https://github.com/apache/tvm/pull/7784
[7777]: https://github.com/apache/tvm/pull/7777
[7816]: https://github.com/apache/tvm/pull/7816
[7534]: https://github.com/apache/tvm/pull/7534
[7809]: https://github.com/apache/tvm/pull/7809
[7739]: https://github.com/apache/tvm/pull/7739
[7788]: https://github.com/apache/tvm/pull/7788
[7792]: https://github.com/apache/tvm/pull/7792
[7887]: https://github.com/apache/tvm/pull/7887
[7897]: https://github.com/apache/tvm/pull/7897
[7794_1]: https://github.com/apache/tvm/pull/7794
[7671]: https://github.com/apache/tvm/pull/7671
[7723]: https://github.com/apache/tvm/pull/7723
[7804_2]: https://github.com/apache/tvm/pull/7804
[7863]: https://github.com/apache/tvm/pull/7863
[7891]: https://github.com/apache/tvm/pull/7891
[7813]: https://github.com/apache/tvm/pull/7813
[7772]: https://github.com/apache/tvm/pull/7772
[7864]: https://github.com/apache/tvm/pull/7864
[7854]: https://github.com/apache/tvm/pull/7854
[7624_2]: https://github.com/apache/tvm/pull/7624
[7797_3]: https://github.com/apache/tvm/pull/7797
[7867]: https://github.com/apache/tvm/pull/7867
[7369]: https://github.com/apache/tvm/pull/7369
[7826]: https://github.com/apache/tvm/pull/7826
[7947]: https://github.com/apache/tvm/pull/7947
[7946]: https://github.com/apache/tvm/pull/7946
[7769]: https://github.com/apache/tvm/pull/7769
[7773]: https://github.com/apache/tvm/pull/7773
[7793]: https://github.com/apache/tvm/pull/7793
[7780]: https://github.com/apache/tvm/pull/7780
[7776]: https://github.com/apache/tvm/pull/7776
[7926]: https://github.com/apache/tvm/pull/7926
[7768]: https://github.com/apache/tvm/pull/7768
[7767]: https://github.com/apache/tvm/pull/7767
[7924]: https://github.com/apache/tvm/pull/7924
[7840]: https://github.com/apache/tvm/pull/7840
[7893]: https://github.com/apache/tvm/pull/7893
[6613]: https://github.com/apache/tvm/pull/6613
[7869]: https://github.com/apache/tvm/pull/7869
[7824]: https://github.com/apache/tvm/pull/7824
[7913]: https://github.com/apache/tvm/pull/7913
[7910]: https://github.com/apache/tvm/pull/7910
[7871]: https://github.com/apache/tvm/pull/7871
[7918]: https://github.com/apache/tvm/pull/7918
[7889]: https://github.com/apache/tvm/pull/7889
[7874]: https://github.com/apache/tvm/pull/7874
[7828]: https://github.com/apache/tvm/pull/7828
[7838]: https://github.com/apache/tvm/pull/7838
[7786_1]: https://github.com/apache/tvm/pull/7786
[7915]: https://github.com/apache/tvm/pull/7915
[7774]: https://github.com/apache/tvm/pull/7774
[7841]: https://github.com/apache/tvm/pull/7841
[7778]: https://github.com/apache/tvm/pull/7778
[7801]: https://github.com/apache/tvm/pull/7801
[7810]: https://github.com/apache/tvm/pull/7810
[7940]: https://github.com/apache/tvm/pull/7940
[7850]: https://github.com/apache/tvm/pull/7850
[7649]: https://github.com/apache/tvm/pull/7649
[7862]: https://github.com/apache/tvm/pull/7862
[7868]: https://github.com/apache/tvm/pull/7868
[7853]: https://github.com/apache/tvm/pull/7853
[7448]: https://github.com/apache/tvm/pull/7448
[7791]: https://github.com/apache/tvm/pull/7791
[7898]: https://github.com/apache/tvm/pull/7898
[7692]: https://github.com/apache/tvm/pull/7692
[7911]: https://github.com/apache/tvm/pull/7911
[7920]: https://github.com/apache/tvm/pull/7920
[7779]: https://github.com/apache/tvm/pull/7779
[7943]: https://github.com/apache/tvm/pull/7943
[7890]: https://github.com/apache/tvm/pull/7890
[7783]: https://github.com/apache/tvm/pull/7783
[7851]: https://github.com/apache/tvm/pull/7851
[7789]: https://github.com/apache/tvm/pull/7789
[7846]: https://github.com/apache/tvm/pull/7846
[7907]: https://github.com/apache/tvm/pull/7907
[7845]: https://github.com/apache/tvm/pull/7845
[7808]: https://github.com/apache/tvm/pull/7808
[7830]: https://github.com/apache/tvm/pull/7830
[7880]: https://github.com/apache/tvm/pull/7880
[7795]: https://github.com/apache/tvm/pull/7795