platform/upstream/tvm.git
4 years ago[Relay] GradientCell Relay Pass (#5039)
Andrew Liu [Tue, 24 Mar 2020 06:56:28 +0000 (23:56 -0700)]
[Relay] GradientCell Relay Pass (#5039)

* save

* gradient.rly

* fix

* NOT WORKING: gradient cell pass

* test gradient pass

* fixed basic call ops

* more tests

* fix bug

* transform calls to one ones_like zero zero_like

* maintenance stuff

* fix linting

* linting

* linting

* throw default

* remove unrelated changes

* import gradent.rly in pass

* comment

* linting

* remove changes to test files

* move gradient_cell.cc to transforms

* revert change

* update files with new commits

* type

* wrapper function to main outermost function type

* fix linting

* fix unsigned and signed int comparison

* review

* GetConstructor definition in module and change op comparison

* update node instantiations

* increase code readability

Co-authored-by: Marisa Kirisame <lolisa@marisa.moe>
4 years agoChange Rust version to stable in Docker (#5138)
Jared Roesch [Tue, 24 Mar 2020 04:05:02 +0000 (21:05 -0700)]
Change Rust version to stable in Docker (#5138)

4 years agoFix for issue #4831. The data_min_idx and data_max_idx were flipped. (#5136)
shoubhik [Tue, 24 Mar 2020 01:33:51 +0000 (18:33 -0700)]
Fix for issue #4831. The data_min_idx and data_max_idx were flipped. (#5136)

Existing test cases cover this fix. In addition I have added an assert to make sure that the data_min is always less than equal to data_max.

4 years agoAdd thrust support for nms (#5116)
Leyuan Wang [Mon, 23 Mar 2020 23:52:33 +0000 (16:52 -0700)]
Add thrust support for nms (#5116)

* add argsort_nms_thrust

* consider valid count in thrust nms sort

* make thrust optional

* typo

* typo

* fix pylint

* address some of the comments

* address more comments

* fix lint

* address more comments

* address more comments

4 years ago[Refactor] Relay Node::make to constructor (#5128)
Zhi [Mon, 23 Mar 2020 18:15:07 +0000 (11:15 -0700)]
[Refactor] Relay Node::make to constructor (#5128)

* relay Node::make to constructor

* patternwildcard

* Address comments

4 years ago[DOCS] Minimize necessary doc change (#5129)
Tianqi Chen [Mon, 23 Mar 2020 17:44:00 +0000 (10:44 -0700)]
[DOCS] Minimize necessary doc change (#5129)

4 years ago[Frontend][TensorFlow]TensorFlow Parser Control Flow Enhancement (#5020)
Yao Wang [Mon, 23 Mar 2020 17:28:19 +0000 (10:28 -0700)]
[Frontend][TensorFlow]TensorFlow Parser Control Flow Enhancement (#5020)

* Improve TF control flow major logic

* Pass mod into operator convert function

* Fix LoopBound

* Add more control flow tests

* Add two test cases for stridedslice

* Fix docstring

* Fix lint

* Fix import

* Fix test assert

* Minor fix conv3d

* Add more comments

* Fix for dilation2d

* Change newly added atan

* Change newly added unravel

4 years ago[Bugfix] Fixed bug where shifting by out-of-bounds value results in no compute code...
pankratz [Mon, 23 Mar 2020 15:47:29 +0000 (09:47 -0600)]
[Bugfix] Fixed bug where shifting by out-of-bounds value results in no compute code being emitted. (#5115)

* Fixed bug where shifting by out-of-bounds RHS values results in LLVM to codegen nothing. Added regression testcase

* Updated testcase to be more precise.

* Fixed testcase

4 years ago[Relay, Topi, TF Frontend] Isfinite operator (#4981)
Mahesh Ambule [Mon, 23 Mar 2020 11:17:12 +0000 (16:47 +0530)]
[Relay, Topi, TF Frontend] Isfinite operator (#4981)

* isfinite doc update

* isfinit expr

* isfinit expr

* isfinite schedule reg

* isfinite python binding

* isfinite python binding

* relay register isfinite

* isfinite type relation

* intrin isfinite

* topi isfinite

* testcase topi isfinite

* tf frontend isfinite

* tf frontend isfinite testcase

* test case relay isfinite

* small fixes

* test forward tf isfinite

* test cases injective for cuda

* remove float16 test case

* add support for isinf

* remove unwanted import

* fix conflict

4 years ago[Relay, Topi] [TF, MXNet] Unravel Index operator (#5082)
Mahesh Ambule [Mon, 23 Mar 2020 01:40:54 +0000 (07:10 +0530)]
[Relay, Topi] [TF, MXNet] Unravel Index operator (#5082)

* first cut unravel_index

* merge fixes

* change rates to dilations

* unravel_index op relay, topi, mxnet, tf

* doc changes

* small changes

* remove empty unravel and argwhere attrs

* remove empty unravel and argwhere attrs

4 years ago[DOCS] Cleanup docs before rebuild (#5127)
Tianqi Chen [Mon, 23 Mar 2020 01:15:19 +0000 (18:15 -0700)]
[DOCS] Cleanup docs before rebuild (#5127)

* [DOCS] Cleanup docs before rebuild

* Ask doxygen to generate svg to minimize the file size

4 years ago[DOC][TUTORIAL] Fix typo for deploy_model_on_android.py (#5123)
Neo Chien [Mon, 23 Mar 2020 01:05:19 +0000 (09:05 +0800)]
[DOC][TUTORIAL] Fix typo for deploy_model_on_android.py (#5123)

4 years ago[CodeGen][CUDA] Vectorization for intrinsics (#5101)
Wei Pan [Sun, 22 Mar 2020 19:22:29 +0000 (12:22 -0700)]
[CodeGen][CUDA] Vectorization for intrinsics (#5101)

- This allows to emit vectorized loads/stores
  for CUDA math intrinsics.

- A few intrinsics should be lowered as CUDAMath not CUDAFastMath ones.

- Fixed the code block identation.

4 years agoAdjust strategy plevel to achieve expected performance by default (#5118)
Haichen Shen [Sun, 22 Mar 2020 16:22:37 +0000 (09:22 -0700)]
Adjust strategy plevel to achieve expected performance by default (#5118)

4 years ago[Rust] Fix the existing test cases before refactoring. (#5122)
Jared Roesch [Sun, 22 Mar 2020 16:22:18 +0000 (09:22 -0700)]
[Rust] Fix the existing test cases before refactoring.  (#5122)

* Fix up the final pieces

* Tweak build.rs

4 years agoUpdate the tarball deployment. (#5120)
Tianqi Chen [Sun, 22 Mar 2020 16:08:46 +0000 (09:08 -0700)]
Update the tarball deployment. (#5120)

4 years ago[DOCS] include a tarball of docs, add a security faq (#5119)
Tianqi Chen [Sun, 22 Mar 2020 03:27:13 +0000 (20:27 -0700)]
[DOCS] include a tarball of docs, add a security faq (#5119)

* [DOCS] include a tarball of docs during deployment

* [DOCS] Add a short security faq

4 years ago[KERAS] conv3d frontend operator support (#5080)
Samuel [Sat, 21 Mar 2020 13:18:41 +0000 (18:48 +0530)]
[KERAS] conv3d frontend operator support (#5080)

* [KERAS]Conv3d support added

* Keras conv3d testcase added

4 years ago[docs] Update relay docs (#5112)
Zhi [Sat, 21 Mar 2020 02:58:48 +0000 (19:58 -0700)]
[docs] Update relay docs (#5112)

* Update relay docs

* any -> py:func

* make clean

4 years agoTVM android camera demo (#5005)
Yagna Srinath Reddy Battula [Sat, 21 Mar 2020 00:29:39 +0000 (17:29 -0700)]
TVM android camera demo (#5005)

4 years ago[Relay][TF] Support for Atan/Atan2 in Relay Tensorflow frontend converter. (#5104)
Andrew Liu [Fri, 20 Mar 2020 23:08:22 +0000 (16:08 -0700)]
[Relay][TF] Support for Atan/Atan2 in Relay Tensorflow frontend converter. (#5104)

* add Atan/Atan2 op

* fix bug and testing

4 years ago[TOPI, Relay refactor] Move Dilation2d from nn to image namespace (#5110)
masahi [Fri, 20 Mar 2020 22:20:13 +0000 (07:20 +0900)]
[TOPI, Relay refactor] Move Dilation2d from nn to image namespace (#5110)

4 years ago[Relay][BYOCG] Propagate constant to subgraphs (#5094)
Zhi [Fri, 20 Mar 2020 21:51:09 +0000 (14:51 -0700)]
[Relay][BYOCG] Propagate constant to subgraphs (#5094)

* bind constant to subgraphs

* con -> constant

4 years ago[Fix] Fix CompilerAttrs (#5109)
Zhi [Fri, 20 Mar 2020 21:29:27 +0000 (14:29 -0700)]
[Fix] Fix CompilerAttrs (#5109)

* fix CompilerAttrs

* retrigger ci

4 years ago[TOPI][OP] Use Thrust sort for argsort and topk (#5097)
MORITA Kazutaka [Fri, 20 Mar 2020 17:00:23 +0000 (02:00 +0900)]
[TOPI][OP] Use Thrust sort for argsort and topk (#5097)

* [TOPI][OP] Use Thrust sort for argsort and topk

The current GPU sort implementation (odd-even transposition sort) is too slow
when the number of elements is large.  This PR introduces Thrust implementation
of sort which is much faster.

Note that this change requires CMake 3.8 or later since we have to use nvcc to
compile a thrust code.

* cmake: make CUDA optional

* allow .cu file to be into the repository

* pylint fix and cleanup

* require cmake 3.8 only when thrust is enabled

* fix nvcc compiler error when passing -pthread

* add missing include

* add USE_THRUST option in config.cmake

* retrigger CI

* retrigger CI

4 years ago[AutoTVM] Temporary fix to the stack overflow issue in autotvm task extraction (...
Haichen Shen [Fri, 20 Mar 2020 16:26:40 +0000 (09:26 -0700)]
[AutoTVM] Temporary fix to the stack overflow issue in autotvm task extraction (#5019)

* Temporary fix to the stack overflow issue in autotvm task extraction

* fix lint

* fix graph tuner test

4 years agoAdd colors to compute_at edges and thread/block indices. (#5111)
yongfeng-nv [Fri, 20 Mar 2020 16:21:45 +0000 (12:21 -0400)]
Add colors to compute_at edges and thread/block indices. (#5111)

4 years ago[TIR][TARGET] Refactor Target codegen to use IRModule and PrimFunc. (#5107)
Tianqi Chen [Fri, 20 Mar 2020 04:02:50 +0000 (21:02 -0700)]
[TIR][TARGET] Refactor Target codegen to use IRModule and PrimFunc. (#5107)

As part of the unified IR refactor.
This PR refactors the target codegen to use IRModule containing tir::PrimFuncs.

In order to break the refactor into several steps without breaking the codebase,
we built an conversion pass to convert Array<LoweredFunc> into IRModule.

The follow-up refactors will gradually move the passes covered by IRModule up
until we cover all the passes. Then we can remove the additional redundant
concepts such as LoweredFunc.

4 years ago[Torch] Add initial 3D op support and test on Resnet 3D (#5075)
masahi [Thu, 19 Mar 2020 20:39:04 +0000 (05:39 +0900)]
[Torch] Add initial 3D op support and test on Resnet 3D (#5075)

* fix minor lint issue

* add conv3d and adaptive avg pool3d conversion with test

* fix max pool handling

* add batch norm 3d test

* add resnet 3d test

* add more conv3d test

* clean up batch norm test

* add note on disabling inception v3 test

* add more tests

* add more tests

* fix names

4 years ago[DOC] Add doc for Relay op strategy (#5078)
Haichen Shen [Thu, 19 Mar 2020 19:22:34 +0000 (12:22 -0700)]
[DOC] Add doc for Relay op strategy (#5078)

* [DOC] Add doc for Relay op strategy

* update

* address more comments

* update

* update

4 years ago[Relay][Frontend][ONNX] operator support NonZero (#5073)
Neo Chien [Thu, 19 Mar 2020 16:49:07 +0000 (00:49 +0800)]
[Relay][Frontend][ONNX] operator support NonZero (#5073)

* [Relay][Frontend][ONNX] operator support: NonZero

* update

* Solve the build fail

* solve the build fail

* Replace ctx_list with tvm.cpu()

4 years ago[ConvertLayout] Support QNN ops. (#5066)
Animesh Jain [Thu, 19 Mar 2020 03:03:56 +0000 (20:03 -0700)]
[ConvertLayout] Support QNN ops. (#5066)

* [ConvertLayout] Support QNN ops.

* Changing layouts to C.

* Fixing dilation.

* Empty commit.

Co-authored-by: Ubuntu <ubuntu@ip-172-31-53-55.us-west-2.compute.internal>
4 years ago[TUTORIAL] Fix vta tutorial after relay function refactor (#5095)
Tianqi Chen [Wed, 18 Mar 2020 23:39:18 +0000 (16:39 -0700)]
[TUTORIAL] Fix vta tutorial after relay function refactor (#5095)

4 years ago[Torch, QNN] Add missing upcast to uint8 avg_pool conversion (#5089)
masahi [Wed, 18 Mar 2020 17:31:06 +0000 (02:31 +0900)]
[Torch, QNN] Add missing upcast to uint8 avg_pool conversion  (#5089)

* add missing upcast to avgpool

* add avg pool test

4 years ago[RELAY] Codegen_c.h should include relay.function (#5093)
lhutton1 [Wed, 18 Mar 2020 17:21:05 +0000 (17:21 +0000)]
[RELAY] Codegen_c.h should include relay.function (#5093)

Change-Id: I015b2c66a50b64d0eb2e9efe336f6c18ea1fdc67

4 years agoDescription updated for pooling attributes (#5091)
Samuel [Wed, 18 Mar 2020 16:06:00 +0000 (21:36 +0530)]
Description updated for pooling attributes (#5091)

4 years ago[CODEGEN][OPENCL] Explicitly cast min/max operands (#5090)
MORITA Kazutaka [Wed, 18 Mar 2020 16:05:49 +0000 (01:05 +0900)]
[CODEGEN][OPENCL] Explicitly cast min/max operands (#5090)

* [CODEGEN][OPENCL] Explicitly cast min/max operands

* retrigger CI

4 years agocreate function.py (#5087)
Zhi [Wed, 18 Mar 2020 15:56:55 +0000 (08:56 -0700)]
create function.py (#5087)

4 years agoReplace UseDefaultCompiler with GetAttr (#5088)
Zhi [Wed, 18 Mar 2020 03:01:15 +0000 (20:01 -0700)]
Replace UseDefaultCompiler with GetAttr (#5088)

4 years agoChange Azure pipeline badge to GH actions (#5081)
Tianqi Chen [Tue, 17 Mar 2020 20:23:43 +0000 (13:23 -0700)]
Change Azure pipeline badge to GH actions (#5081)

4 years ago[RELAY][PY] Fix relay node registration after refactor (#5083)
Tianqi Chen [Tue, 17 Mar 2020 17:59:23 +0000 (10:59 -0700)]
[RELAY][PY] Fix relay node registration after refactor (#5083)

4 years ago[Refactor][Relay] Refactor Relay Python to use new FFI (#5077)
Zhi [Tue, 17 Mar 2020 15:33:10 +0000 (08:33 -0700)]
[Refactor][Relay] Refactor Relay Python to use new FFI (#5077)

* refactor relay python

* revert relay/ir/*.py to relay

* Address comments

* remove direct access to analysis and transform namespace

4 years ago[Relay, TF Frontend] Dilation2D operator support (#5033)
Mahesh Ambule [Tue, 17 Mar 2020 06:54:32 +0000 (12:24 +0530)]
[Relay, TF Frontend] Dilation2D operator support (#5033)

* update docs for dilation 2d

* dilation2d compute

* dilation2d register

* dilation2d rel compute

* dilation2d strategy

* dilation2d attrs

* dilation2d generic schedule

* dilation2d tf frontend support

* dilation2d tf frontend test case

* dilation2d test cases

* pylint fixes

* add exception for cuda target

* Update docstring

* Update docstring

* change rates to dilations

* removed unused param

* merge master

* Update nn.py

* Update nn.py

4 years ago[TFLITE]DepthToSpace and SpaceToDepth support (#5041)
Samuel [Mon, 16 Mar 2020 19:57:06 +0000 (01:27 +0530)]
[TFLITE]DepthToSpace and SpaceToDepth support (#5041)

* [TFLITE]DepthToSpace and SpaceToDepth op parser support

* DepthToSpace and SpaceToDepth testcases

* Review comments fixed

4 years agoTensorflow script upgrade from 1.13.1 to 2.0.0, so that it can run in both versionsw...
Samuel [Mon, 16 Mar 2020 19:55:48 +0000 (01:25 +0530)]
Tensorflow script upgrade from 1.13.1 to 2.0.0, so that it can run in both versionsw (#4963)

4 years agoReturn empty CSourceModule when no lowered_funcs exists in Relay mod (#4847)
Ruizhe Zhao [Mon, 16 Mar 2020 19:52:45 +0000 (19:52 +0000)]
Return empty CSourceModule when no lowered_funcs exists in Relay mod (#4847)

* Use dummy func when no lowered_funcs exists in Relay mod

* Dummy func -> CSourceModule with empty code str

* Added comments describing the empty CSouceModule

* Always import external modules w/o assertions

* Use CSourceModule as a fallback for LLVMModule

* Changed cond for target == llvm

* Create an empty LLVM module w/o using dummy func

* Avoid using IR str concat to create LLVM module

* Improved comments for codegen.LLVMModuleCreate

* Satisfy the linter for LLVMModuleCreate

4 years ago[Team] jwfromm -> reviewer (#5076)
Yizhi Liu [Mon, 16 Mar 2020 10:01:54 +0000 (03:01 -0700)]
[Team] jwfromm -> reviewer (#5076)

4 years ago[IR] Update the type_keys to reflect the code-org (#5074)
Tianqi Chen [Sun, 15 Mar 2020 22:39:47 +0000 (15:39 -0700)]
[IR] Update the type_keys to reflect the code-org (#5074)

4 years ago[Relay, TOPI] Refactor Adaptive pool and add 3d support (#5049)
masahi [Sun, 15 Mar 2020 20:43:48 +0000 (05:43 +0900)]
[Relay, TOPI] Refactor Adaptive pool and add 3d support (#5049)

* add stub for nd impl

* refactored indices compute

* refactored divide step

* remove unused variables, add doc

* fix lint

* add relay op def

* add python registration

* refactor topi test

* update relay tests, but test result is weird

* workaround for weird bug

* add relay adaptive pool 3d test

* add topi tests

* update doc for 3d

* typo fix

* fix lint

* add more tests including NDHWC

4 years agoAdd support for FusedBatchNormV3 (#5065)
ojotoxy [Sun, 15 Mar 2020 19:02:59 +0000 (12:02 -0700)]
Add support for FusedBatchNormV3 (#5065)

No changes seem to be needed to _fused_batch_norm. It just works.

4 years ago[TESTS] Triage the testcases to fit the the new namespaces (#5071)
Tianqi Chen [Sun, 15 Mar 2020 04:40:50 +0000 (21:40 -0700)]
[TESTS] Triage the testcases to fit the the new namespaces (#5071)

* [TESTS] Triage the testcases to fit the naming convention of the new namespaces

* Remove multiple usage of system lib to avoid test problems

4 years ago[TIR] Introduce tir::PrimFunc (#5070)
Tianqi Chen [Sat, 14 Mar 2020 21:26:23 +0000 (14:26 -0700)]
[TIR] Introduce tir::PrimFunc (#5070)

This PR introduces tir::PrimFunc which will be used as the TIR function
container in the unified IR.

Also streamlined the function attributes a bit further.
- All common attributes are under tvm::attr
- TIR specific attributes are under tvm::tir::attr and comes with a tir prefix
- Use stl_style for attributes for now

4 years ago[COMMUNITY] Add @abergeron -> reviewer (#5064)
Jared Roesch [Sat, 14 Mar 2020 04:02:03 +0000 (21:02 -0700)]
[COMMUNITY] Add @abergeron -> reviewer (#5064)

4 years ago[Torch, QNN] Remove FP32 piggy back and use QNN add/mul/concatenate (#5061)
masahi [Fri, 13 Mar 2020 22:11:32 +0000 (07:11 +0900)]
[Torch, QNN] Remove FP32 piggy back and use QNN add/mul/concatenate (#5061)

* use qnn add/mul/concatenate

* remove logging

4 years ago[Relay][AutoTVM] Bug Fix for ARM CPUs. Lower strict assumption. (#5063)
Bernhard [Fri, 13 Mar 2020 21:44:14 +0000 (22:44 +0100)]
[Relay][AutoTVM] Bug Fix for ARM CPUs. Lower strict assumption. (#5063)

4 years ago[Relay][Pass] Add submodule extraction pass (#4960)
anwang2009 [Fri, 13 Mar 2020 20:58:49 +0000 (13:58 -0700)]
[Relay][Pass] Add submodule extraction pass (#4960)

* rebased

* fix lint

4 years ago[Graph tuner]Add opt out operator for has_multiple_inputs for graph tuner (#5000)
zhen-jia [Fri, 13 Mar 2020 17:07:06 +0000 (10:07 -0700)]
[Graph tuner]Add opt out operator for has_multiple_inputs for graph tuner (#5000)

* consider layout_transform in has_multiple_inputs

* refactor code

* remove debug info

* remove subclass assignment

* refactoring a little bit

* remove default value

* remove trailing whitespace

* modify test for has_multiple_inputs

Co-authored-by: Ubuntu <ubuntu@ip-172-31-40-194.us-west-2.compute.internal>
4 years ago[DOCS] Move git_howto to rst, add Stage documents to te (#5055)
Tianqi Chen [Fri, 13 Mar 2020 03:23:56 +0000 (20:23 -0700)]
[DOCS] Move git_howto to rst, add Stage documents to te (#5055)

4 years ago[PY] Require python3.6 (#5057)
Tianqi Chen [Fri, 13 Mar 2020 03:22:38 +0000 (20:22 -0700)]
[PY] Require python3.6 (#5057)

4 years agomaintenance (#5058)
Andrew Liu [Fri, 13 Mar 2020 00:22:19 +0000 (17:22 -0700)]
maintenance (#5058)

4 years ago[Bugfix][IR][ATTRS] Fix AttrEqual for Array and StrMap, double (#5054)
Tianqi Chen [Thu, 12 Mar 2020 23:29:06 +0000 (16:29 -0700)]
[Bugfix][IR][ATTRS] Fix AttrEqual for Array and StrMap, double (#5054)

- Use fuzzy comparison for double.
- Removed the hack for BatchNormAttrs and DictAttr.

Also removed a warning from text printer printing.

4 years ago[C++] Require c++14 by default (#5056)
Tianqi Chen [Thu, 12 Mar 2020 22:31:56 +0000 (15:31 -0700)]
[C++] Require c++14 by default (#5056)

4 years agoCI: Install apt-transport-https (#5053)
Marcus Shawcroft [Thu, 12 Mar 2020 17:13:22 +0000 (17:13 +0000)]
CI: Install apt-transport-https (#5053)

The ubuntu_install_llvm.sh script started failing because of a http to
https redirect.  This patch adds the package that allows apt to handle
https transport.

Change-Id: I70bcba32a9fc75d02c54f4f21f288b2f46226689

4 years ago[Autotvm] Fix autotvm customized template (#5034)
Haichen Shen [Thu, 12 Mar 2020 17:01:22 +0000 (10:01 -0700)]
[Autotvm] Fix autotvm customized template (#5034)

* init

* fix template

* tweak naming

4 years ago[Strategy] Support for Int8 schedules - CUDA/x86 (#5031)
Animesh Jain [Thu, 12 Mar 2020 16:51:09 +0000 (09:51 -0700)]
[Strategy] Support for Int8 schedules - CUDA/x86 (#5031)

* [CUDA] Op strategy changes for Int8 schedules.

* Applying Haichen's suggestions.

* Make 4D output work for task extraction.

* Make x86 work.

* Fix lint.

* Lint fixes.

* Tests, comments, out channel a multiple of 4.

* Topi test.

Co-authored-by: Ubuntu <ubuntu@ip-172-31-38-96.us-west-2.compute.internal>
4 years ago[1] Test case modified for int type (#5012)
ANSHUMAN TRIPATHY [Thu, 12 Mar 2020 16:50:13 +0000 (22:20 +0530)]
[1] Test case modified for int type (#5012)

4 years agoFixed div by zero core dump. Fixed rounding intrinsics on int crash (#5026)
pankratz [Thu, 12 Mar 2020 16:35:36 +0000 (10:35 -0600)]
Fixed div by zero core dump. Fixed rounding intrinsics on int crash (#5026)

4 years ago[REFACTOR] Streamline Function Attr interface. (#5045)
Tianqi Chen [Thu, 12 Mar 2020 15:33:47 +0000 (08:33 -0700)]
[REFACTOR] Streamline Function Attr interface. (#5045)

* [REFACTOR] Streamline Function Attr interface.

There has been quite a few recent changes that depends heavily on
the function attr interface. This PR streamlines that interface by introducing
two APIs that covers most of the usages.

- GetAttr which gets a typed object for a given key
  - HasNonzeroAttr is a quick helper that calls GetAttr to quickly check an attribute
- WithAttr that creates a new function object with the given attr
  - The API comes with copy on write optimization to avoid multiple copies
  - We deliberately pick the prefix With(instead of Set) to indicate this
    function does not mutate the original input.

On the python side:
- We allow read access via func.attrs (which is a DictAttr)
- func.with_attrs to create a new instance with updated attrs.

We also get rid of the small wrapper functions and make sure the API centered around
the GetAttr and HasNonzeroAttr interface.

This PR also changes the function construction to follow the new convention.

* Address review comments

* Address review comments

* Fix doxygen path

4 years ago[TFLITE][FRONTEND]Reduce_any op parsing support (#4926)
Samuel [Thu, 12 Mar 2020 07:11:37 +0000 (12:41 +0530)]
[TFLITE][FRONTEND]Reduce_any op parsing support (#4926)

* [TFLITE][FRONTEND]Reduce_any op parsing support

* Testcase check added to run in tf version above 1.14.0 & review comments

* Review comment, checked updated to 1.15

4 years ago[TFLITE]Round op parsing support added (#5022)
Samuel [Thu, 12 Mar 2020 07:09:45 +0000 (12:39 +0530)]
[TFLITE]Round op parsing support added (#5022)

4 years agoSupport for AddV2 in Relay Tensorflow frontend converter. (#5046)
Fernand Pajot [Thu, 12 Mar 2020 07:09:00 +0000 (01:09 -0600)]
Support for AddV2 in Relay Tensorflow frontend converter. (#5046)

4 years agoSet split node's range to minimum of ext and split factor or split nparts, but only...
yongfeng-nv [Thu, 12 Mar 2020 03:59:23 +0000 (23:59 -0400)]
Set split node's range to minimum of ext and split factor or split nparts, but only when PassDownDomain is called with allow_missing == false, i.e. by InferBound. Add a helper PassUpThreadBinding() to get a map telling whether an IterVar has at least one leaf IterVar deriving from it binding to a thread. Add two unit tests. (#5044)

4 years ago[VTA] VTA hardware/software codebase re-org (#5037)
Thierry Moreau [Thu, 12 Mar 2020 03:59:03 +0000 (20:59 -0700)]
[VTA] VTA hardware/software codebase re-org (#5037)

4 years ago[refactor][relay pass] Separate analysis and transform passes (#5035)
Zhi [Thu, 12 Mar 2020 03:47:21 +0000 (20:47 -0700)]
[refactor][relay pass] Separate analysis and transform passes (#5035)

* [refactor][relay pass] Separate analysis and transform passes into different subfolders

* remove pass folder

4 years ago[Object] Add String container (#4628)
Wei Chen [Wed, 11 Mar 2020 22:53:01 +0000 (06:53 +0800)]
[Object] Add String container (#4628)

4 years agoConditions updated to cover better user scenarios[Re-raised] (#5043)
ANSHUMAN TRIPATHY [Wed, 11 Mar 2020 21:36:39 +0000 (03:06 +0530)]
Conditions updated to cover better user scenarios[Re-raised] (#5043)

* Conditions updated to cover better user scenarios

* [1] New test case added

* [2] New test case added

* [3] Proper variable name used

* [4] Review Comments handled

* [5] Review comments handled

* [6] Review comments handled

4 years ago[Relay][VM] Fix compilation of If-Elses (#5040)
Wei Chen [Wed, 11 Mar 2020 19:26:28 +0000 (03:26 +0800)]
[Relay][VM] Fix compilation of If-Elses (#5040)

4 years agoConv3D ONNX support and conv3D_ncdhw x86 schedules (#4949)
Matthew Brookhart [Wed, 11 Mar 2020 16:03:30 +0000 (09:03 -0700)]
Conv3D ONNX support and conv3D_ncdhw x86 schedules (#4949)

* Support 3d Convolution with the ONNX frontend

* add unit tests for conv3d in onnx frontend

respond to PR formatting requests

add x86 schedules to conv3d ncdhw test

fix a doc string format issue

refactor for changed upsream API

* first attempt at conv3d autotuning

add default schedule for conv3d_ncdhw

fill in autotvm integration

add a fallback for invalid schedules

fix fallback

fix reduction order to get simd working correctly

4 years ago[Intrin] Adding a few missing math intrin (#5011)
Bing Xu [Wed, 11 Mar 2020 15:58:20 +0000 (08:58 -0700)]
[Intrin] Adding a few missing math intrin  (#5011)

* [intrin] exp2

* [intrin] exp10

* [intrin] log2/10

* [intrins] exp10

* [test] math intrin

4 years agoRevert "Tighten split's extent (#4931)" (#5027)
Lianmin Zheng [Wed, 11 Mar 2020 15:57:30 +0000 (08:57 -0700)]
Revert "Tighten split's extent (#4931)" (#5027)

This reverts commit 585f9ce6e7bef7d0e8902b1c1e55dcb3bbe84eed.

4 years agoRevert "Conditions updated to cover better user scenarios (#4951)" (#5032)
Tianqi Chen [Wed, 11 Mar 2020 15:56:03 +0000 (08:56 -0700)]
Revert "Conditions updated to cover better user scenarios (#4951)" (#5032)

This reverts commit fe74b37ab578e6d3c540b0f6ac187a220ccc028a.

4 years ago[QNN] Support 4D padding. (#5036)
Animesh Jain [Wed, 11 Mar 2020 15:30:13 +0000 (08:30 -0700)]
[QNN] Support 4D padding. (#5036)

* [QNN] Support 4D padding.

* Empty commit.

Co-authored-by: Ubuntu <ubuntu@ip-172-31-38-96.us-west-2.compute.internal>
4 years ago[TFLITE]Activation functions support (#4978)
Samuel [Wed, 11 Mar 2020 03:08:48 +0000 (08:38 +0530)]
[TFLITE]Activation functions support (#4978)

* [TFLITE]elu, leaky_relu, lrn, log_softmax activation functions

* removed ops present in pr 4805

* review_comments updated

4 years ago[CodeGen][CUDA] Enhance CUDA codegen for SelectNode (#4983)
Wei Pan [Wed, 11 Mar 2020 02:16:59 +0000 (19:16 -0700)]
[CodeGen][CUDA] Enhance CUDA codegen for SelectNode (#4983)

- This patch allows CUDA backend to emit correct code for
  selects with vector conditions, which may be produced
  by floordiv op lowering etc..

- This already works for llvm BE, as llvm select instruction
  supports vector conditions.

Signed-off-by: Wei Pan <weip@nvidia.com>
4 years ago[topi][relay] new PR to re-add tan to TVM (#5025)
notoraptor [Wed, 11 Mar 2020 01:24:04 +0000 (21:24 -0400)]
[topi][relay] new PR to re-add tan to TVM (#5025)

* Add relay operation relay.op.tan.

* Update tan implementation in TVM.

* Update tests.

* Add shape function for tan.

* Add missing main test to python/frontend/tensorflow/test_forward.

* Revert, back to sin/cos.

* Revert "Revert, back to sin/cos."

This reverts commit 4da5b503b921585ba9d80944b29136142b575c40.

* Fix implementation of tan in cuda. Do not support tan for float16.

Simplify topi/tests/python/test_topi_math. Add testing for tan with float32 and float64.

Finally implement tan as sin/cos in llvm.

4 years ago[CI] Temporary disable rust test (#5029)
Tianqi Chen [Tue, 10 Mar 2020 18:46:56 +0000 (11:46 -0700)]
[CI] Temporary disable rust test (#5029)

4 years ago[Torch] Add initial control flow support (#4964)
masahi [Tue, 10 Mar 2020 17:59:09 +0000 (02:59 +0900)]
[Torch] Add initial control flow support  (#4964)

* Add support for prim::If and prim::Loop with test cases

* rebase and fix tests

* add some comments

* simplifying, fix float cast

* parse -> convert

* recursivly retrive ops in get_all_op_names

* use multiple return values from block correctly, simplify loop convert

* choose dtype properly for zeros and ones

* simplifying, replace convert_inputs with _get_relay_input_vars

* fix for while loop with non input dependent init cond

* add assert on loop var update

* move the condition around

* better testing for seg models

* rebase fix, disable inception v3 in quant test as it is too slow to
load with torch-1.4 + torchvision 0.5

* simplify and add more comparison op converter

4 years agoRevert "[topi][relay] add operation tan to TVM (#4938)" (#5017)
Yao Wang [Tue, 10 Mar 2020 09:37:12 +0000 (02:37 -0700)]
Revert "[topi][relay] add operation tan to TVM (#4938)" (#5017)

This reverts commit d992468d80af816f0413fc43c2ee1c02f7fe19c3.

4 years ago[RELAY] Remove primitive attribute from composite function (#5014)
lhutton1 [Tue, 10 Mar 2020 08:10:07 +0000 (08:10 +0000)]
[RELAY] Remove primitive attribute from composite function (#5014)

* A composite function should not be primitive since we still may need to perform passes on it.

Change-Id: If62d06d265234861a6ec0df7749dc1c339c1055c

4 years agoEarly checking added and new test cases added for schedule fuse (#5010)
ANSHUMAN TRIPATHY [Tue, 10 Mar 2020 04:23:18 +0000 (09:53 +0530)]
Early checking added and new test cases added for schedule fuse (#5010)

* [1] New test case added for fuse

* [2] New test case added for fuse

* [3] New test case added for fuse

* [4] New test case added for fuse

* [5] Early check added

4 years agoImplemented kDLCPUPinned (cudaMallocHost) (#4985)
jmorrill [Tue, 10 Mar 2020 02:39:41 +0000 (19:39 -0700)]
Implemented kDLCPUPinned (cudaMallocHost) (#4985)

* implement kDLCPUPinned

* Fix line endings

* Fix whitespace for linter

* cleanup up allocdataspace method

4 years agoRevive the Rust + SGX refactor (#4976)
Jared Roesch [Tue, 10 Mar 2020 02:35:20 +0000 (19:35 -0700)]
Revive the Rust + SGX refactor (#4976)

* Add Nick's changes's squashed

* Fix frontend compilation

* Re-enable Rust CI

* Add changes with conflicted badly

* Restructure import_module! macro in order to avoid unstable features

* Kill old unstable feature enablement

* Refactor common to use new APIs

* Move the code to stable

* Fix warning

Co-authored-by: Nick Hynes <nhynes@oasislabs.com>
4 years ago[REDO AFTER GH BUG] Add support for quantized models via QNN (#5016)
masahi [Tue, 10 Mar 2020 01:36:50 +0000 (10:36 +0900)]
[REDO AFTER GH BUG] Add support for quantized models via QNN (#5016)

This reverts commit f346c60287b50950275e20db9e6d84b3fc568a00.

4 years agoRevert "[Torch, QNN] Add support for quantized models via QNN (#4977)" (#5013)
Animesh Jain [Mon, 9 Mar 2020 20:14:58 +0000 (13:14 -0700)]
Revert "[Torch, QNN] Add support for quantized models via QNN (#4977)" (#5013)

This reverts commit fc7f0783940c362bf48cd46817956381196201e2.

4 years agotypo (#5008)
雾雨魔理沙 [Mon, 9 Mar 2020 19:50:23 +0000 (12:50 -0700)]
typo (#5008)

4 years ago[Runtime] MISRA-C compliant TVM runtime (#3934)
Liangfu Chen [Mon, 9 Mar 2020 18:21:28 +0000 (02:21 +0800)]
[Runtime] MISRA-C compliant TVM runtime (#3934)

* implement of MISRA-C compliant TVM runtime;

* working on bundle_deploy_c demo

* move header files into include dir

* fix compatibility issues

* fix compatibility issues

* resolve most of the warnings and errros

* implement c_backend_api

* introduce bridge

* working well

* move to header files and bundle.c into src/runtime/crt

* clean up

* satisfy linter

* clean up

* test with the cat image

* remove synset

* refactoring

* refactoring

* refactoring

* initial crt_runtime_api.c

* improved compatibility with g++

* using exposed API in c_runtime_api.h

* call from c_runtime_api.h

* clean up

* lint

* merge into apps/bundle_deploy directory

Change-Id: I51904db81b8589e65d107d8ca77b47452e3812b5

* make the demo runs in ci

Change-Id: I2c24f8b592508833d3555311c2b24d1931f19385

* address review comments

Change-Id: I027ddff15c31fb4da0bd0e461427dce619de1f93

* release

Change-Id: I5ad5bb8426468aac9fc8d074e56ddea358a7fd91

* fix ci testing

Change-Id: Ic2e82fb3051b6c254ef32a964f976b61e3e5fe4d

* add test case for misra c runtime

Change-Id: Ie0dfd0ade6be4665b4384db7d260a6c69b35010f

* fread files in testing to avoid calling xxd

Change-Id: Ie7fbc16b4b0b9509918d986a841f443900813bef

4 years ago[VTA][Chisel,de10nano] Chisel fixes and de10nano support (#4986)
Pasquale Cocchini [Mon, 9 Mar 2020 17:09:13 +0000 (10:09 -0700)]
[VTA][Chisel,de10nano] Chisel fixes and de10nano support (#4986)

* [VTA][de10nano] Enable user defined target frequency.

Issue:
The VTA target frequency on the DE10-Nano is hardcoded to 50MHz
unnecessarily limiting performance.

Solution:
Add a PLL to the FPGA sub-system along with support for the
selection of a user specified frequency at build time. The board
successfully builds and runs at 100MHz.

* Added a PLL in the soc_system.tcl platform designer generator
  script.

* Modified the Makefile to automatically set the target frequency
  from that specified in the pkg_config.py file.

* Modified the Makefile to generate a bitstream with an RBF
  format that enables programming of the FPGA directly from
  the on-board processor. Specifically, the RBF is generated in
  FastParallel32 mode with compression, which corresponds to the
  default MSEL switch setting on the board, i.e. 01010.

* Added a false path override to file set_clocks.sdc to turn off
  unconstrained path warnings on the VTA pulse LED.

* [VTA][TSIM] Add more debug and tracing options.

* Modified Makefile to change default config to DafaultDe10Config.

* Added option in Makefile to produce more detailed tracing
  for extra observability in debugging complex scenarios.

* Added option in Makefile to produce traces in FST format which
  are 2 orders of magnitude smaller, although much slower to
  generate.

* Added option in Makefile to build the simulator with GCC address
  sanitizer.

* Modified Makefile to not lint the scala code by default avoiding
  unintended wrong indentation. Linting should be better performed
  manually on a per-need basis.

* [VTA][de10nano] Enable remote programming of FPGA.

Issue:
The Cyclone V FPGA on board of the DE10-Nano can only be programmed
using the JTAG port, which is a limiting option for users.

Solution:
Add support for the remote programming of the FPGA implementing
the FPGA programming manager protocol published in the Cyclone V
user manual.

* Added file de10nano_mgr.h implementing an FPGA manager class
  that supports handling of control and status registers as well
  as a push-button option to program the FPGA. The class can be
  easily extended to include more registers if needed.

* Used an instance of the FPGA manager to implement function
  VTAProgram also warning users when incompatible bitstream
  files are used.

* Registered VTAProgram as a global function and modified
  the program_bitstream python class to use it.

* [VTA][de10nano] Enhance de10nano runtime support.

Issue:
The de10nano target has incomplete, non-working support
for runtime reconfiguration, bitstream programming, and
examples of usage.

Solution:
Complete runtime support for the de10nano target.

* Modified VTA.cmake to comment out a default override for
  VTA_MAX_XFER to 21 bit wide.

* Modified VTA.cmake to add needed de10nano include dirs.

* Modified relevant files to support de10nano same way as
  other targets for VTA runtime reconfiguration and FPGA
  programming.

* Added test_program_rpc.py example as a runtime FPGA
  programming example. Note that unlike the pynq target
  no bitstream is either downloaded or programmed when
  the bitstream argument is set to None.

* Cosmetic changes to vta config files.

* [VTA][Chisel] LoadUop FSM bug fix.

Issue:
The LoadUop FSM incorrectly advances the address of the next
uop to read from DRAM when the DRAM data valid bit is deasserted
and asserted at the end of a read. This is caused by a mismatch
in the logic of the state and output portions of the FSM.
This is one of two issues that was gating the correct operation
of VTA on the DE10-Nano target.

Solution:
Modify the logic of the output section of the FSM to include
a check on the DRAM read valid bit or fold the output assignemnt
into the state section.

* Folded the assignemnt of the next uop address in the state
  section of the FSM.

* [VTA][Chisel] Dynamically adjust DMA tranfer size.

Issue:
In the DE10-Nano target and possibly in others, DMA transfers that
cross the boundaries of memory pages result in incorrect reads and
writes from and to DRAM. When this happens depending on different
input values, VTA loads and stores exhibit incorrect results for
DMA pulses at the end of a transfer. This is one of two issues that
were gating the DE10-Nano target from functioning correctly, but may
affect other Chisel based targets.

Solution:
Add support for dynamically adjustble DMA transfer sizes in load
and store operations. For a more elegant and modular implementation
the feature can be enabled at compile time with a static constant
that can be passed as a configuration option.

* Modified the load and store finite state machines to dynamically
  adjust the size of initial and stride DMA transfers. The feature
  is enabled by default by virtue of the static constant
  ADAPTIVE_DMA_XFER_ENABLE.

* [VTA][Chisel] Improve FSIM/TSIM/FPGA xref debug.

Issue:
Cross reference between FSIM, TSIM, and Chisel based FPGA traces
is an invaluable instrument that enables fast analysis on FSIM,
and analysis/debug on TSIM and FPGA, especially for complex flows
like conv2d or full inferences. Currently this cannot be done
easily since a suitable reference is missing. The clock cycle
event counter cannot be used since it is undefined in FSIM and
not reliable between TSIM and FPGA because of different latencies.

Solution:
Introduce a new event counter that preserves a program order across
FSIM, TSIM, FPGA. We propose adding the accumulator write event
counter in the Chisel EventCounter class and a simple instrumentation
in the FSIM runtime code. Note that this technique enabled finding the
Chisel issues reportes in the PR, which would have been otherwise
far more difficult.

* Added the acc_wr_count event counter and changed interfaces
  accordingly.

* [VTA][de10nano] Comply with linting rules.

* [VTA] Appease make lint.

* [VTA] Disable pylint import not top level error.

* [VTA][Chisel,de10nano] Linting changes.

* Use CamelCase class names.

* Use C++ style C include header files.

* Add comments to Chisel makefile.

* [VTA][de10nano]

* Reorder C and C++ includes in de10nano_mgr.h.

* Restore lint as default target in Chisel Makefile.

* [VTA][de10nano] Do not use f string in pkg_config.py.

* [VTA][de10nano] Remove overlooked f strings in pkg_config.py.

* [VTA][de10nano] Fixed typo.

* [VTA][TSIM] Check if gcc has align-new.

* [VTA][Chisel] Make adaptive DMA transfer default.

* [VTA][RPC] Renamed VTA_PYNQ_RPC_* to VTA_RPC_*.

Issue:
With more FPGA targets coming online the initial method of
using individual environment variables to specify target IP and port
does not scale well.

Solution:
Use a single VTA_RPC_HOST, VTA_RPC_PORT pair to be changed
every time a different target is used. For instance in a script
used to benchmark all targets.

* Replaced every instance of VTA_PYNQ_RPC_HOST and VTA_PYNQ_RPC_PORT
  with VTA_RPC_HOST and VTA_RPC_PORT, respectively.

* [VTA][Chisel] Comply with new linter.

4 years agoDocs and Readme updated as per new namespace change (#4989)
ANSHUMAN TRIPATHY [Sun, 8 Mar 2020 16:01:18 +0000 (21:31 +0530)]
Docs and Readme updated as per new namespace change (#4989)

4 years agolower plevel of conv2d winograd on cuda (#4987)
Haichen Shen [Sun, 8 Mar 2020 11:51:02 +0000 (04:51 -0700)]
lower plevel of conv2d winograd on cuda (#4987)

4 years agokill from tvm import te (#5007)
Michal Jamroz [Sun, 8 Mar 2020 06:31:11 +0000 (07:31 +0100)]
kill from tvm import te (#5007)

Co-authored-by: Michal Jamroz <jamroz@chem.uw.edu.pl>