TVM Monthly - June 2021

不念不忘少年蓝@ 2022-09-09 14:59 90阅读 0赞

# [TVM Monthly - June 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.

# Features and Improvements Highlights #

*  Meta Schedule (AutoTIR), the new auto-scheduling system on top of TensorIR is under the RFC process ([https://github.com/apache/tvm-rfcs/pull/5 14][https_github.com_apache_tvm-rfcs_pull_5_14])
 *  The first PR of Automatic Mixed Precision Pass is merged ([https://github.com/apache/tvm/pull/8069 6][https_github.com_apache_tvm_pull_8069_6])
 *  Unified Static Memory Planning is under the RFC process ([https://github.com/apache/tvm-rfcs/pull/9 2][https_github.com_apache_tvm-rfcs_pull_9_2])
 *  The first PR of texture support is merged ([https://github.com/apache/tvm/pull/7711 1][https_github.com_apache_tvm_pull_7711_1])

# Community #

In June of 2021, we welcomed many new contributors to the project. Importantly, we welcome [@echuraev][echuraev], [@wyc-ruiker][wyc-ruiker] as the new reviewers. Thanks to everyone for the hard work and contributions!

This forum got 115k pageviews, 2.7k user visits in the last month.

# Pull Requests #

The below is a high-level summary of the PRs closed in the last month grouped by area.

# Relay #

*  Update unique operator to match ONNX output (1D only) [\#8099][8099]
 *  Refactor strided\_slice and add axes argument [\#8165][8165]
 *  Support converting TF combined\_nms using Relay all\_class\_nms [\#8174 2][8174_2]
 *  Convert a fake quantized or QAT graph into QNN ops [\#8126 1][8126_1]
 *  make simplify inference iterative [\#8246 2][8246_2]
 *  Enable layout transformation for image.resize op [\#8205][8205]
 *  Additional gradients [\#8307][8307]
 *  Enable group\_conv1d import through conv2d conversion. [\#8321][8321]
 *  fix first-order AD tuple/projection expr duplication [\#8318][8318]
 *  Add negative log likelihood loss (nll\_loss) op [\#8056][8056]
 *  Remove in-place modification of attributes in layout transform [\#8309 1][8309_1]
 *  Fix test\_rewrite\_function\_with\_fuzzy\_body test check [\#8287][8287]
 *  Fix type relation for batch\_matmul [\#8376 1][8376_1]
 *  Support slash in identifier. [\#8352][8352]
 *  Resize 1D [\#8346][8346]
 *  Fix index order in conv2d computation for Arm CPU. [\#8361][8361]
 *  Add support of conv2d with NHWC for Mali [\#8422][8422]
 *  Add support of conv2d with NHWC for Bifrost [\#8430][8430]
 *  Add mixed precision (e.g. FP16) model conversion pass [\#8069 6][https_github.com_apache_tvm_pull_8069_6]
 *  Modify create\_executor to pass params [\#8418][8418]
 *  Batch\_matmul to dense optimization [\#8440 2][8440_2]
 *  Add ConvInteger support. [\#8456 2][8456_2]
 *  Add RandomUniform converter and tests to onnx frontend. [\#8426][8426]
 *  Allow importing models with malformed Loop nodes. [\#8475][8475]
 *  Switch from CompileEngine to TECompiler in Interpreter [\#8486][8486]

# Vm #

*  Add function attribute for shape func for profiling [\#8148][8148]
 *  Improve error messages [\#8218][8218]
 *  try reallocation once when OOM [\#8285][8285]
 *  Fix the shape function of conv & Add dynamic support for conv2d nhwc [\#8480 1][8480_1]
 *  Allow serialization of function attrs which are strings [\#8485][8485]

# Vulkan #

*  Minor updates following Vulkan target query implementation [\#8151][8151]
 *  Remove dependency on Target from -from\_device functionality. [\#8171][8171]
 *  Compatibility fix for test\_vulkan\_unique() [\#8186][8186]
 *  Corrected typo in Vulkan capability error messages [\#8187][8187]
 *  Pull out vulkan initialization into VulkanInstance and VulkanDevice [\#8188][8188]
 *  Move ownership of per-CPU-thread objects to VulkanDeviceAPI [\#8196][8196]
 *  Fixed SPIR-V scoping bug with threadIdx [\#8281][8281]
 *  Implement sync for SyncThread("warp") [\#8320][8320]
 *  Added debug saving of Vulkan shaders, environment variable documentation [\#8333][8333]
 *  Improved error message for extern calls passed to SPIR-V codegen [\#8332][8332]
 *  Add parametrization to vulkan unit tests. [\#8348][8348]

# Tvmc #

*  Fix tvmc compile to extract target and target\_host from --target [\#8176][8176]
 *  Add a --config option to `tvmc compile` [\#8253][8253]
 *  Fix inconsistent usage of host\_name -> hostname [\#8324][8324]
 *  Add vulkan to targets of tvmc run. [\#8359][8359]

# Ci #

*  Fix Lint Issue [\#8185][8185]
 *  Fix the sphinx doc style for sphinx4 [\#8198][8198]
 *  Fix python file format due to changes in black [\#8194][8194]
 *  Use pre-built binaries instead of compiled [\#8245][8245]
 *  Install curl in the context of ubuntu\_install\_nodejs.sh [\#8326][8326]
 *  Pin mypy version [\#8329][8329]

# Metal #

*  Fix the rest memory leaks in Metal runtime [\#8175][8175]
 *  Remove matching Metal to OpenCL in tophub [\#8211][8211]
 *  Reduce number of threads for reduction layers [\#8206][8206]
 *  Fix run metal model when non first device is selected [\#8261][8261]
 *  Fix bad stream after interrupted tuning session [\#8244][8244]
 *  Add pass for splitting kernel with huge number of args [\#8313][8313]

# Refactor #

*  Re-organize Containers into SubFolders [\#8183][8183]
 *  Remove dead code from depthwise\_conv2d for Intel graphics [\#8381][8381]
 *  Enforce attaching storage scope to PointerType [\#8366][8366]

# Topi #

*  Fix conv2d\_nchw for opencl intel graphics [\#8201][8201]
 *  Allow cblas batch\_matmul implicit batch\_size broadcast [\#8250][8250]
 *  Removed requirement for GPU from topi conv2d\_cudnn.cuda and conv3d\_cudnn.cuda [\#8276 1][8276_1]
 *  Parametrized tests in `test_topi_dense.py`, split out gpu-independent implementations [\#8336][8336]
 *  Bugfix for topi.prod [\#8416][8416]
 *  Fix the integer overflow problem of the scatter\_nd op. [\#8415][8415]
 *  Add support for arbitrary dtypes to CSRMV and CSRMM [\#8437][8437]

# Fix #

*  Allow tokenizer to parse numbers greater than INT\_MAX. [\#8120][8120]
 *  QEMU RPC issue [\#8021][8021]
 *  Detect like cores by looking at scaling\_max\_freq instead of cpuinfo\_max\_freq [\#8370][8370]
 *  Fix depthwise conv2d on non-cuda GPU platforms [\#8379][8379]
 *  Fix wrong log of tir pass VerifyMemory [\#8445][8445]
 *  Explicitly retain `__hash__` of `StringImm` [\#8449][8449]
 *  Update stale relay.Module API in docs/comments [\#8411 1][8411_1]
 *  Remove unused variable in GraphExecutorCodegen [\#8465][8465]
 *  Compiler supports input with a slash [\#8481][8481]
 *  Minor misspelling [\#8476 1][8476_1]

# Frontend #

*  Adding TF2 frontend code with support for control flow ops [\#8142][8142]
 *  Speed up Fold Constant [\#8208 1][8208_1]
 *  Tensorflow frontend op refactor [\#8179 1][8179_1]
 *  Support for broadcasting in batch\_matmul when shapes differ [\#8251][8251]
 *  Adding InversePermutation Op [\#8277][8277]
 *  Check LLVM enabled/installed [\#8414 1][8414_1]
 *  Stridedslice and concat\_v2 fix [\#8483][8483]

# Tir #

*  Fix data dependent indexing when lowering TE to TIR [\#8217][8217]
 *  Tighten up invariance of CopyOnWrite in recursive stmt visitor [\#8358][8358]
 *  fix storage rewrite index remap [\#8338][8338]

# Tvmscript #

*  Add tir.min node in tvm script [\#8219][8219]
 *  Fix printing of rank 0 buffer access [\#8215][8215]
 *  add more type support in script function parameter [\#8235][8235]
 *  specialize [\#8354][8354]
 *  TVMScript Parser support BufferSlice indices [\#8408][8408]
 *  Support tir.abs node in tvm script [\#8488][8488]

# Microtvm #

*  make RVM memory and number of cores variable [\#8154][8154]
 *  Add QEMU build to RVM image [\#8190][8190]
 *  Add wait to QEMU Setup [\#8236][8236]
 *  Add Nucleo stm32l4r5zi board to zephyr [\#8386][8386]
 *  Refactor uTVM to microTVM [\#8283][8283]
 *  Add fixture to zephyr test [\#8393][8393]
 *  Fix Stack Size Issue for Zephyr AOT Demo on Physical Hardware [\#8453][8453]
 *  Fix clock skew on virtualbox [\#8395][8395]

# Docker #

*  Add external directory mount [\#8144][8144]
 *  Update tensorflow/tflite/xgboost versions [\#8306][8306]
 *  fix sphinx install versions [\#8316][8316]
 *  Fix ordering of tf and tflite installs in ci\_qemu [\#8315][8315]
 *  Update lint to reflect the latest state [\#8330][8330]
 *  Update gpg server [\#8319][8319]
 *  fix keras install [\#8391][8391]

# Unittests #

*  Automatic parametrization over targets, with explicit opt-out [\#8010][8010]
 *  Added a meta-test for tvm.testing.fixture behavior in case of a broken fixture. [\#8343][8343]
 *  Minor fixes to unit tests for cudnn/vulkan targets [\#8462][8462]

# Docs #

*  Prevented docs/1 file from being generated. [\#8029][8029]
 *  Fix for broken link in apps for wasm-standalone dir [\#8045][8045]
 *  Add docs for Pass Instrument [\#8220][8220]
 *  Corrected typo in googletest build instructions. [\#8459][8459]

# Onnx #

*  Support Bidirectional RNNs [\#8337][8337]
 *  Wrap 'If' if it has multiple outputs [\#8385][8385]

# Bugfix #

*  fix segfault when cudnnDestroy called with destroyed cuda context [\#8267][8267]
 *  Fix broadcast type func with incomplete type [\#8438][8438]
 *  do not simplify 'Any() - Any()' to 0 [\#8266][8266]

[TVM Monthly - June 2021]: https://discuss.tvm.apache.org/t/tvm-monthly-june-2021/10546
[https_github.com_apache_tvm-rfcs_pull_5_14]: https://github.com/apache/tvm-rfcs/pull/5
[https_github.com_apache_tvm_pull_8069_6]: https://github.com/apache/tvm/pull/8069
[https_github.com_apache_tvm-rfcs_pull_9_2]: https://github.com/apache/tvm-rfcs/pull/9
[https_github.com_apache_tvm_pull_7711_1]: https://github.com/apache/tvm/pull/7711
[echuraev]: https://github.com/echuraev
[wyc-ruiker]: https://github.com/wyc-ruiker
[8099]: https://github.com/apache/tvm/pull/8099
[8165]: https://github.com/apache/tvm/pull/8165
[8174_2]: https://github.com/apache/tvm/pull/8174
[8126_1]: https://github.com/apache/tvm/pull/8126
[8246_2]: https://github.com/apache/tvm/pull/8246
[8205]: https://github.com/apache/tvm/pull/8205
[8307]: https://github.com/apache/tvm/pull/8307
[8321]: https://github.com/apache/tvm/pull/8321
[8318]: https://github.com/apache/tvm/pull/8318
[8056]: https://github.com/apache/tvm/pull/8056
[8309_1]: https://github.com/apache/tvm/pull/8309
[8287]: https://github.com/apache/tvm/pull/8287
[8376_1]: https://github.com/apache/tvm/pull/8376
[8352]: https://github.com/apache/tvm/pull/8352
[8346]: https://github.com/apache/tvm/pull/8346
[8361]: https://github.com/apache/tvm/pull/8361
[8422]: https://github.com/apache/tvm/pull/8422
[8430]: https://github.com/apache/tvm/pull/8430
[8418]: https://github.com/apache/tvm/pull/8418
[8440_2]: https://github.com/apache/tvm/pull/8440
[8456_2]: https://github.com/apache/tvm/pull/8456
[8426]: https://github.com/apache/tvm/pull/8426
[8475]: https://github.com/apache/tvm/pull/8475
[8486]: https://github.com/apache/tvm/pull/8486
[8148]: https://github.com/apache/tvm/pull/8148
[8218]: https://github.com/apache/tvm/pull/8218
[8285]: https://github.com/apache/tvm/pull/8285
[8480_1]: https://github.com/apache/tvm/pull/8480
[8485]: https://github.com/apache/tvm/pull/8485
[8151]: https://github.com/apache/tvm/pull/8151
[8171]: https://github.com/apache/tvm/pull/8171
[8186]: https://github.com/apache/tvm/pull/8186
[8187]: https://github.com/apache/tvm/pull/8187
[8188]: https://github.com/apache/tvm/pull/8188
[8196]: https://github.com/apache/tvm/pull/8196
[8281]: https://github.com/apache/tvm/pull/8281
[8320]: https://github.com/apache/tvm/pull/8320
[8333]: https://github.com/apache/tvm/pull/8333
[8332]: https://github.com/apache/tvm/pull/8332
[8348]: https://github.com/apache/tvm/pull/8348
[8176]: https://github.com/apache/tvm/pull/8176
[8253]: https://github.com/apache/tvm/pull/8253
[8324]: https://github.com/apache/tvm/pull/8324
[8359]: https://github.com/apache/tvm/pull/8359
[8185]: https://github.com/apache/tvm/pull/8185
[8198]: https://github.com/apache/tvm/pull/8198
[8194]: https://github.com/apache/tvm/pull/8194
[8245]: https://github.com/apache/tvm/pull/8245
[8326]: https://github.com/apache/tvm/pull/8326
[8329]: https://github.com/apache/tvm/pull/8329
[8175]: https://github.com/apache/tvm/pull/8175
[8211]: https://github.com/apache/tvm/pull/8211
[8206]: https://github.com/apache/tvm/pull/8206
[8261]: https://github.com/apache/tvm/pull/8261
[8244]: https://github.com/apache/tvm/pull/8244
[8313]: https://github.com/apache/tvm/pull/8313
[8183]: https://github.com/apache/tvm/pull/8183
[8381]: https://github.com/apache/tvm/pull/8381
[8366]: https://github.com/apache/tvm/pull/8366
[8201]: https://github.com/apache/tvm/pull/8201
[8250]: https://github.com/apache/tvm/pull/8250
[8276_1]: https://github.com/apache/tvm/pull/8276
[8336]: https://github.com/apache/tvm/pull/8336
[8416]: https://github.com/apache/tvm/pull/8416
[8415]: https://github.com/apache/tvm/pull/8415
[8437]: https://github.com/apache/tvm/pull/8437
[8120]: https://github.com/apache/tvm/pull/8120
[8021]: https://github.com/apache/tvm/pull/8021
[8370]: https://github.com/apache/tvm/pull/8370
[8379]: https://github.com/apache/tvm/pull/8379
[8445]: https://github.com/apache/tvm/pull/8445
[8449]: https://github.com/apache/tvm/pull/8449
[8411_1]: https://github.com/apache/tvm/pull/8411
[8465]: https://github.com/apache/tvm/pull/8465
[8481]: https://github.com/apache/tvm/pull/8481
[8476_1]: https://github.com/apache/tvm/pull/8476
[8142]: https://github.com/apache/tvm/pull/8142
[8208_1]: https://github.com/apache/tvm/pull/8208
[8179_1]: https://github.com/apache/tvm/pull/8179
[8251]: https://github.com/apache/tvm/pull/8251
[8277]: https://github.com/apache/tvm/pull/8277
[8414_1]: https://github.com/apache/tvm/pull/8414
[8483]: https://github.com/apache/tvm/pull/8483
[8217]: https://github.com/apache/tvm/pull/8217
[8358]: https://github.com/apache/tvm/pull/8358
[8338]: https://github.com/apache/tvm/pull/8338
[8219]: https://github.com/apache/tvm/pull/8219
[8215]: https://github.com/apache/tvm/pull/8215
[8235]: https://github.com/apache/tvm/pull/8235
[8354]: https://github.com/apache/tvm/pull/8354
[8408]: https://github.com/apache/tvm/pull/8408
[8488]: https://github.com/apache/tvm/pull/8488
[8154]: https://github.com/apache/tvm/pull/8154
[8190]: https://github.com/apache/tvm/pull/8190
[8236]: https://github.com/apache/tvm/pull/8236
[8386]: https://github.com/apache/tvm/pull/8386
[8283]: https://github.com/apache/tvm/pull/8283
[8393]: https://github.com/apache/tvm/pull/8393
[8453]: https://github.com/apache/tvm/pull/8453
[8395]: https://github.com/apache/tvm/pull/8395
[8144]: https://github.com/apache/tvm/pull/8144
[8306]: https://github.com/apache/tvm/pull/8306
[8316]: https://github.com/apache/tvm/pull/8316
[8315]: https://github.com/apache/tvm/pull/8315
[8330]: https://github.com/apache/tvm/pull/8330
[8319]: https://github.com/apache/tvm/pull/8319
[8391]: https://github.com/apache/tvm/pull/8391
[8010]: https://github.com/apache/tvm/pull/8010
[8343]: https://github.com/apache/tvm/pull/8343
[8462]: https://github.com/apache/tvm/pull/8462
[8029]: https://github.com/apache/tvm/pull/8029
[8045]: https://github.com/apache/tvm/pull/8045
[8220]: https://github.com/apache/tvm/pull/8220
[8459]: https://github.com/apache/tvm/pull/8459
[8337]: https://github.com/apache/tvm/pull/8337
[8385]: https://github.com/apache/tvm/pull/8385
[8267]: https://github.com/apache/tvm/pull/8267
[8438]: https://github.com/apache/tvm/pull/8438
[8266]: https://github.com/apache/tvm/pull/8266