* Add shape inference for Ops used by BERT
* Erf
* Pow
* ReduceMean
* Dropout
* Expand
https://github.com/onnx/onnx/blob/master/docs/Operators.md#expand
Deduce the value of the shape operand by looking at the producer
of the operand.
Currently supported producers are: onnx.Constant and onnx.Shape.
* Add corresponding tests for each op.
* Sort the list of ops with shape inference in gen_onnx_mlir.py
in alphabetic order for clarity.
* Restart CI
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* base implementation
* add example
* change table gen
* docs
* small change for review
Co-authored-by: Alexandre Eichenberger <alexe@us.ibm.com>
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* move scalerop to decompose
* change clang format
* change clang format
* add shape inference for scaler op
* fixing generated onnxop
* generate onnx.md
* add benefit for scaler decompose and simplify scaler shape inference
* cast rewrite only for float
* add cast op same type rewrite rule
* fix format
Co-authored-by: chentong319 <chentong@us.ibm.com>
* Define krnl.permute op.
* Support krnl.permute operation.
* Properly remove loop references.
* Re-push, Github was down.
* Need to debug interpretOp error.
* Fix lowering bug by erasing ops after full krnl IR interpretation is done, and clean up & comment code.
* Introduce permute, unroll operations.
* More debug.
* Remove std::set.
* krnl.terminate fails to be converted.
* Pass all tests, need to add legal ops as well as part of the conversion target.
* Change test format to new permute spec.
* Bug fix for nested iterate op lowering.
* Simplify error reporting.
* Fix compilation error.
* Increase comments coverage.
* Remove unnecessary imports.
* Re-trigger Jenkins
* Add permute/unroll tests.
* Retrigger Jenkins
* Using a non-trivial example.
* Add more complex example/test case.
* move scalerop to decompose
* change clang format
* change clang format
* add shape inference for scaler op
* fixing generated onnxop
* generate onnx.md
* Add shape inference for scaler op
* add benefit for scaler decompose and simplify scaler shape inference
* Reorganize main function.
* Follow review comments.
* Emit constants are globals in Krnl and LLVM dialects.
* Add support for moving dynamic alloca instructions to top of functions.
* Fix memory pooling tests.
* Various fixes.
* Fix lit tests.
* More test fixes.
* Reformat.
* Reformat some more.
* Fix issue with TestConv and split-input-file.
* Use smart pointers.
* Remove redundant pointer.
* Reformat.
* Add initMap description.
* Clean up tests.
* Remove optimize_loops/return_loops op in elementwise ops lowering and fix tests in onnx_lowering.mlir.
* Fix all tests.
* Remove all occurences of def_loops/return_loops.
* Fix test.
* Fix comments for defineLoops & emitKrnlLoopsAndIterationForOperand function.
* Remove emitOptimizedLoops.
* Allow not specifying optimizedLoops when creating KrnlIterateOperandPack.
* Fix style.
* Make BuildKernelLoop helper not emit optimize/return_loop operations & retire emitKrnlLoopsAndIterationForOperand by replacing it with BuildKernelLoop.
* DefineLoops -> DefineLoopsEx, remove redundant emitKrnlLoopsAndIterationForOperand function.
* BuildKrnlLoop API name update.
* Tweak comments.
* Remove unused withEmptyOptimization flag.
* Better comment for BuildKrnlLoop.
* Fully remove krnl.return_loops/optimize_loops op.
* Trigger Windows Build
* Bump windows ci python version.
* Move to more recent LLVM ID (May 15)
* clang-format
* Bump cache version up
* Update readme
* Fix doc check
* Move to a newer commit id
* Update LoopToStandard -> SCFToStandard
* Change MLIRSideEffects to MLIRSideEffectInterfaces
* Add AffineScope trait to KrnlIterateOp
* [ElementWise] Load/Store op to AffineLoad/AffineStore op
* [Gemm, MatMul, Reduction, Softmax] Load/Store op to AffineLoad/AffineStore op
* [Concat] Load/Store op to AffineLoad/AffineStore op
* [Pad, PadConstantValuePad, Reshape, Transpose] Load/Store op to AffineLoad/AffineStore op
* [LSTM] Load/Store op to AffineLoad/AffineStore op
* [Conv, Norm, Pooling] Load/Store op to AffineLoad/AffineStore op
* Add affine-loop-fusion pass
* Use Load/Store for scalar
* Use Load/Store for scalar
* Fix lit tests
* Unknown dimensions for broadcasting ops
* Affine Load/Store for scalar memref
* clang-format
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* Lower Squeeze op to Krnl dialect
* Emit tensor size as a single constant; add a lit test for unknown dimensions
* Code style
* Speical case where the input is only used by this squeeze op
* Remove squeeze-in-place optimization
* Update ConvertONNXToKrnl.cpp
Twek to re-run tests.
* Trigger buildbot re-run.
* Re-run CI
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* Support krnl.block printing/parsing.
* Checkpoing, PoC working.
* Implement krnl.block operation.
* Make tuple -> make pair.
* Bug fix, white list krnl.iterate op while lowering.
* Add return loop op lowering.
* Bug fix.
* Allow using loop refs more than once if they are used by krnl.iterate op.
* More comments and include lit test.
* Make krnl.block definition more restrictive.
* Splitting tests creates modules, making affine_map matching more verbose, prefer not splitting since test cases are small.
* Use verbose mode for LIT test on Z.
* Use verbose build to diagnose.
* Missing libraries linking when building in shared mode.
* Fix whole-archive linkage.
* Try preloading affinetransforms.
* Try put AffineTransforms into LD_LIBRARY_PATH.
* Fix python syntax error.
* No need to link with whole-archive libs, as they are pre-loaded.
* Do not preload any library.
* Link with whole-archive libs.
* Explicitly shared linkage in CMake.
* Fix CMake syntax error.
* Restore test.py
* Update z13.sh
* Update z13.sh
* Provide krnl.block operation description.
* string type from tensorflow
* simplify type
* parser and print
* gen StringType for tablegen
* onnx to onnx-mlir type
* add namespace
* allow all integer type
* dialect document
* add test case
* format
* more precise type for ONNXOp
* format
* enable the failed test
* update comment
* update onnx.md
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
* Explicit pass registration.
* Remove whole-archive linking, replace with regular linking.
* Remove whole-archive linkage related scripts.
* No need to preload library, simply expose them through LD_LIBRARY_PATH.
* Use OMLibs to record all onnx-mlir libs.
* Add OMResultTypeInferenceOpInterface lib to OMLibs.
* nit.
* No need to expose libs through LD_LIBRARY_PATH.
* Fix missing onnx header file issue.
* Define OMLibs before Tool subdirectory is imported.
* Define OMLibs at parent scope.
* Specify dependency of MainUtils on OMLibs early.
* Set OMLibs both at current & parent scope.
* Add comment about what future pass implementation should do.
* support map and seq in tablegen
* register MLONNX for testing
* format
* remove the unwanted test
* add a test
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* initial const prop attempt
* added support for broadcast ops
* adde all binary broadcast ops into custom builders with precise type
* added test example
* working
* format
* fixed suggestion by Tung, start woring on unary
* added subtraction and neg the right way, and added elementwise mul too
* formatting changes
* format
* format
* added instructions to add new optimizations
* added propagation rules that always migrate constants toward the root of the expression, using assoc and commutativity
* format comment
* PoC works.
* MNist works.
* Clean up.
* Fix test.
* Make Linux work.
* Use consistent symbol name.
* Fix variable name.
* Fix array addr access.
* Bug fix.
* Bug fix.
* install before running e2e tests.
* Fix build config.
* Use sudo when installing.
* Make embeddedDataLoader position independent.
* Enable ResNet50.
* Format code.
* Format MainUtil.
* Try not using sudo to install.
* Supply runtime dir via environment variable.
* Dump problematic operation.
* Dump entire function.
* Debug.
* Dump input.
* Dump constant op.
* Debug.
* Debug.
* Debug.
* Print to stderr.
* take care of endianness.
* Use endianness-aware execution session.
* Fix ZLinux error.
* Include warning when desired output endianness can't be deduced.
* Remove debug code.
* Remove debug code in shape inference.
* Support binary-decoder for testing constants packing.
* Support filename, move-to-file, elision-threshold configurations in constant packing pass for easy testing.
* Add lit test, fix lit test type mismatch.
* Add more consts packing tests.
* Ensure intermediate files are properly cleaned up.
* No need for constant elimination.
* Link with threading libraries.
* Remove debug code.
* Format code.
* More tests.
* test nit.
* Remove debug code.
* Reduce hard-coded constants.
* Use temporary and unique working directory for hosting model parameters.
* Test if it works.
* Try to find objcopy.
* Rename symbols using objcopy.
* Move sanitized name to linux section.
* Use verbose mode for debugging.
* Disambiguate pass constructor.
* Fix symbol name.
* Use Command API to build and execute commands.
* Move linux to use Command API.
* Fix reset args.
* Execute redefine sym.
* Format code.
* Do not use verbose mode for CircleCI.
* Remove debug code.
* Prettify code, add comments.
* getSegmentData -> getEmbeddedConstPool
* vector -> std::vector.
* Make sure we properly clean up intermediate files.
* Fix test cases.
* Add runtime directory.
* Trigger rebuild.
* [Merge with master] fix debug script.
* Diable affine fusion pass for now.
* Support generic fallback const packing mechanism.
* Remove debug code.
* Handle the case where objcopy is not available.
* Fix Windows missing types.
* Support int64.
* Copy packed constant to a local directory for non-Linux/Mac platforms.
* Nit: remove debug code, refactor const pack preprocessing out as a separate function.
* Cannot make preprocessConstPack a standalone function because file removers are stack-allocated, and they are deallocated prematurely when function stack gets popped, deleteing intermediate files too early.
* Don't require executable filename.
* Import ONNX data types directly.
* Fix LIT test.
* Bug fix, use moved string value.
* Remove redundant filenames.
* Fix CMake script.
* Embed endianness information as a symbol, and check during runtime.
* More comments, update lit tests.
* Fix lit test on BE machine.
* Copyright notices.
* Reorganize main function.
* Follow review comments.
* Emit constants are globals in Krnl and LLVM dialects.
* Replace internal malloc with memory pool and getref instruction.
* Lower krnl.getref to LLVM.
* Fix formatting issues.
* Add tests.
* Add missing dependency.
* Improve LLVM lowering.
* Add test to show getref is generic.
* 1. Add shape inference for the following ops:
- Atan
- Tan
- Sin
- Cast
- ConvTranspose
- Flatten
- DynamicQuantizeLinear
- QuantizeLinear
- DequantizeLinear
- ConvInteger
2. Import attributes for generic nodes
3. Fixes for cases where .cast<> should be .isa<> (ONNXConcat::inferShapes)
* Fix foormatting issues
* Address comments:
- SmallVector<> * -> SmallVectorImpl<> &
- switch-case -> helper function
- Inside helper function, preserve signed-ness
- add TODOs
* Can't use signed integers yet in convertONNXTypeToMLIRType, add TODO
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* initial const prop attempt
* added support for broadcast ops
* adde all binary broadcast ops into custom builders with precise type
* added test example
* working
* format
* fixed suggestion by Tung, start woring on unary
* added subtraction and neg the right way, and added elementwise mul too
* formatting changes
* format
* format
* added instructions to add new optimizations
* enable promote attr for pad
* use optional arguments for pad
* shape infereance for pad
* Lowering Pad
* format file
* use DenseTensor for the attribute
* use Pad in ONNXRewrite
* fix the merge conflict
* fix the attr given to constantOp
* handle ONNXConstantOp in attribute promotion
* Fix bug when AttributePromotion is called more than once
* update ONNXOps.td.inc with correct version of onnx
* update onnx.md
* responses to review
* fix the build error
* change the implementation of Pad
* delete commented out code
* clang format
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* Use AffineMap
* Shared AffineMap
* AffineMap for Conv/Pooling
* Create helper files
* Remove changes for Relu
* Remove redundant includes
* Use AffineMap for AveragePool's shape inference
* Add MLIR tests for unknown dimension case
* Extract a method AffineMapIntConstant
* Comment stylist and include path
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
* Specialize the op lowering logic for elementwise operations
* Fix clang-format error.
* Update tests for LSTM since LSTM uses element-wise ops
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* Support dilations and enable e2e tests
* Fix allocating memory for dynamic shape
* Edit comments
* Do dilation by computing an offset from kernel index
* Correct dilation formula, add an example of out-of-bound, and add a test for dilation
* Import optional outputs as NoneType
* Shape inference for ONNXLSTM
* Edit ONNXLSTM::inferShape()
* Shape inference for ONNXLSTMOp
* Create a common function for inferring shape for RNN ops
* CheckInsertDeallocation for a specific result
* Allocate memory for LSTM
* First round of lowering
* Allocate memory for hidden and cell states
* Test with custom Tanh
* Fix an error in Ct's formula
* Add E2E tests
* Return outputs
* Refactor the code
* Enable E2E tests
* Support reverse and bidirectional directions
* Minor revision
* Return all intermediate hidden states
* Call existing activation functions
* Structs for activation functions
* Call existing activations in ONNX
* Minor revision
* Compare strings ignoring case
* Use memreftype of rank 0 for calling activation functions
* Fix getActivationPack()
* Revise the code
* Add one MLIR test
* Add MLIR tests for reverse and bidirectional modes
* Make the order of emiting instructions deterministic
* Use OperandAdaptor instead of directly use an operand index
* Use literal assignments
* Change some variable names
* Use literal assignments
* Use literal assignments
* Format the code
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
* Implement shape inference for SplitOp
* Change spitOpt to SplitAttribute and check the axis range before updating the axis attribute
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
* Reorganize main function.
* Follow review comments.
* Emit constants are globals in Krnl and LLVM dialects.
* Output of non-value constants. Write full source to file.
* Fix e2e tests.
* Output constant free and full code in separate files.
* Emit separate files.
* Move file output management to utils.
* Elide the values of glotbal krnl constants.
* Add dual file output for Basic flag.
* Add tests.
* Add passes to cmake file.
* Move to more recent LLVM commit ID
* Update LLVM cache version from V9 to V10
* Update to latest LLVM commit id from master, roll back conditions in util scripts
* Update circlci LLVM cache tag to ensure ci updates builds with latest LLVM commit id
* Update README.md to have matching LLVM commit id
* Update doc/Dialtects/onnx.md
* Enable onnx-mlir for VS builds on Windows
* Update README to include lit
* Update build command for Windows to include config
* Update build instructions, add cmd files for windows, enable single source of truth for MLIR commit-id (clone-mlir.sh)
* Add Visual Studio workload info
* Update ONNX op definitions
* Revert onnx submodule back to previous commit, disable warnings in CMakeLists to work around build issues with MSVC
* Update environment for path to PDcurses on Windows
* Fix directory strings to be compatible with Windows or Linux style slashes
* Fix install-mlir.sh so it works when sourced
* Ensure README and cmd files match and have correct paths
* Properly quote ONNX_MLIR_SRC_DIR
* Address PR feedback: Use llvm_unreachable to indicate failure to convert attribute proto to name/value pair
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* Create a template for pooling and add support for AveragePool
* Edit MLIR tests for MaxPool according to the new lowering template for pooling
* Dealloc temporary variables
* Support count_include_pad for AveragePool
* Add MLIR tests for AveragePool lowering
* Make changes according to Tian's comments
* Push AffineMap as upper bound for KrnlIterateOp
* Test AffineMap to use in Pooling
* Replace the old implementaion by a new one using AffineMap
* Fix the computation when dilations are non-unit
* Clean up the old code
* Remove AveragePool from Canonicalization pass
* Fix computing the end indices of a filter window
* Refactor the code for pooling
* Revise pushAffineMapBound
* Add MLIR tests
* Remove unused functions
* Fix check-onnx-backend build on x86 Linux. (#91)
* Add the split marker to test files (#90)
Co-authored-by: Tian Jin <tjingrant@gmail.com>
Co-authored-by: gongsu832 <gong_su@hotmail.com>
Co-authored-by: Tian Jin <tjingrant@gmail.com>
* Specify each lib only once; allow llvm build in shared libs mode.
* Remove debug code.
* For library targets, retain dependency information using add_dependencies, but do not link using taget_link_libraries.
* Do not set LD_PRELOAD by default.
Co-authored-by: Gong Su <gongsu@us.ibm.com>
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
* implement shape inference for concat
* better checking of axis being concatenated: constant values only
* lowering of Concat with lit and backend tests
* fixes
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
* Bug fix, ensure krnl.iterate can lower in the degenerate case.
* Fix parser issue with degenerate iterate op.
* Add a test case.
* Remove dead code.
Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
albertomagni-ms
Tian Jin
chentong319
Gheorghe-Teodor Bercea
Anh Leu
Tung D. Le
Anh Leu
Aaron Smith
Alexandre Eichenberger
Aman LaChapelle
Byron Changuion