Initialize conversion passes from ONNX to Torch backend and add IR tests for ONNXAddOp and ONNXConstantOp

Signed-off-by: Quinn Dawkins <[email protected]>

Author: qedawkins

.gitmodules

@@ -13,3 +13,7 @@
 [submodule "third_party/mlir-hlo"]
 	path = third_party/mlir-hlo
 	url = https://github.com/tensorflow/mlir-hlo.git
+[submodule "third_party/torch-mlir"]
+	path = third_party/torch-mlir
+	url = https://github.com/llvm/torch-mlir
+	shallow = true

CMakeLists.txt

@@ -8,6 +8,7 @@ project(onnx-mlir)
 option(ONNX_MLIR_BUILD_TESTS "Build ONNX-MLIR test executables. If OFF, just generate build targets." ON)
 option(ONNX_MLIR_CCACHE_BUILD "Set to ON for a ccache enabled build." OFF)
 option(ONNX_MLIR_ENABLE_MHLO "Enable MHLO support." ON)
+option(ONNX_MLIR_ENABLE_TORCH "Enable Torch support." ON)
 option(ONNX_MLIR_ENABLE_WERROR "Enable warnings as errors." OFF)
 option(ONNX_MLIR_SUPPRESS_THIRD_PARTY_WARNINGS "Suppress warning in third_party code." ON)
 
@@ -152,6 +153,12 @@ if (ONNX_MLIR_ENABLE_MHLO)
   add_subdirectory(third_party/mlir-hlo EXCLUDE_FROM_ALL)
 endif()
 
+if (ONNX_MLIR_ENABLE_TORCH)
+  set(TORCH_MLIR_ENABLE_MHLO OFF)
+  set(TORCH_MLIR_OOT_BUILD ON)
+  add_subdirectory(third_party/torch-mlir EXCLUDE_FROM_ALL)
+endif()
+
 if (NOT TARGET benchmark)
   set(BENCHMARK_USE_BUNDLED_GTEST OFF)
   set(BENCHMARK_ENABLE_GTEST_TESTS OFF)
@@ -185,6 +192,10 @@ if (ONNX_MLIR_ENABLE_MHLO)
   add_compile_definitions(ONNX_MLIR_ENABLE_MHLO)
 endif()
 
+if (ONNX_MLIR_ENABLE_TORCH)
+  add_compile_definitions(ONNX_MLIR_ENABLE_TORCH)
+endif()
+
 add_subdirectory(utils)
 add_subdirectory(include)
 add_subdirectory(src)

docs/BuildOnLinuxOSX.md

@@ -26,12 +26,15 @@ cmake -G Ninja ../llvm \
    -DLLVM_TARGETS_TO_BUILD="host" \
    -DCMAKE_BUILD_TYPE=Release \
    -DLLVM_ENABLE_ASSERTIONS=ON \
-   -DLLVM_ENABLE_RTTI=ON
+   -DLLVM_ENABLE_RTTI=ON \
+   -DMLIR_ENABLE_BINDINGS_PYTHON=ON
 
 cmake --build . -- ${MAKEFLAGS}
 cmake --build . --target check-mlir
 ```
 
+If building onnx-mlir with `ONNX_MLIR_ENABLE_TORCH=OFF`, then `MLIR_ENABLE_BINDINGS_PYTHON` is not needed.
+
 ## ONNX-MLIR (this project)
 
 ### Build

src/Conversion/CMakeLists.txt

@@ -9,3 +9,7 @@ add_subdirectory(ONNXToTOSA)
 if (ONNX_MLIR_ENABLE_MHLO)
   add_subdirectory(ONNXToMhlo)
 endif()
+
+if (ONNX_MLIR_ENABLE_TORCH)
+  add_subdirectory(ONNXToTorch)
+endif()

src/Conversion/ONNXToTorch/CMakeLists.txt

@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Please keep in alphabetical order.
+
+install(TARGETS
+  TorchMLIRTorchDialect
+  TorchMLIRTorchUtils
+  )
+
+add_onnx_mlir_library(OMONNXToTorch
+  ConvertONNXToTorch.cpp
+  ConvertONNXToTorchPipeline.cpp
+  EraseONNXEntryPoint.cpp
+  ONNXToTorchCommon.cpp
+  TypeConversion/TorchTypeConversion.cpp
+  TypeConversion/TorchTypeConversionPasses.cpp
+
+  Math/Elementwise.cpp
+
+  Tensor/Constant.cpp
+
+  LINK_LIBS PUBLIC
+  TorchMLIRTorchDialect
+  TorchMLIRTorchUtils
+  OMONNXOps
+  )

src/Conversion/ONNXToTorch/ConvertONNXToTorch.cpp

@@ -0,0 +1,84 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//====------ ConvertONNXToTorch.cpp - ONNX dialects to Torch lowering
+//-------===//
+//
+// Copyright 2019-2022 The IBM Research Authors.
+//
+// ===============================================================================
+//
+// This file implements the lowering of frontend operations to Torch backend IR.
+//
+//===------------------------------------------------------------------------===//
+
+#include "src/Conversion/ONNXToTorch/ONNXToTorchCommon.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+
+void populateONNXToTorchConversionPattern(TypeConverter &typeConverter,
+    RewritePatternSet &patterns, MLIRContext *ctx) {
+  // Math
+  populateLoweringONNXElementwiseOpToTorchPattern(typeConverter, patterns, ctx);
+  populateLoweringONNXConstantOpToTorchPattern(typeConverter, patterns, ctx);
+}
+
+//===----------------------------------------------------------------------===//
+// Frontend to Mhlo Dialect lowering pass
+//===----------------------------------------------------------------------===//
+
+struct FrontendToTorchLoweringPass
+    : public PassWrapper<FrontendToTorchLoweringPass, OperationPass<ModuleOp>> {
+
+  StringRef getArgument() const override { return "convert-onnx-to-torch"; }
+
+  StringRef getDescription() const override {
+    return "Lower frontend ops to Torch dialect.";
+  }
+
+  // Make sure that we have a valid default constructor and copy
+  // constructor to make sure that the options are initialized properly.
+  FrontendToTorchLoweringPass() = default;
+  FrontendToTorchLoweringPass(const FrontendToTorchLoweringPass &pass)
+      : PassWrapper<FrontendToTorchLoweringPass, OperationPass<ModuleOp>>() {}
+
+  void runOnOperation() final;
+};
+
+void FrontendToTorchLoweringPass::runOnOperation() {
+  ModuleOp module = getOperation();
+  // The first thing to define is the conversion target. This will define the
+  // final target for this lowering.
+  ConversionTarget target(getContext());
+
+  // We define the specific operations, or dialects, that are legal targets for
+  // this lowering.
+  target.addLegalDialect<torch::Torch::TorchDialect, func::FuncDialect>();
+
+  TypeConverter typeConverter;
+  typeConverter.addConversion([](Type type) { return type; });
+  onnx_mlir::setupTorchTypeConversion(target, typeConverter);
+
+  // Now that the conversion target has been defined, we just need to provide
+  // the set of patterns that will lower the frontend operations.
+  RewritePatternSet patterns(&getContext());
+
+  // Define patterns.
+  populateONNXToTorchConversionPattern(typeConverter, patterns, &getContext());
+
+  // With the target and rewrite patterns defined, we can now attempt the
+  // conversion. The conversion will signal failure if any of our `illegal`
+  // operations were not converted successfully.
+  if (failed(applyPartialConversion(module, target, std::move(patterns)))) {
+    signalPassFailure();
+  }
+}
+
+std::unique_ptr<Pass> createLowerToTorchPass() {
+  return std::make_unique<FrontendToTorchLoweringPass>();
+}
+
+} // namespace onnx_mlir

src/Conversion/ONNXToTorch/ConvertONNXToTorchPipeline.cpp

@@ -0,0 +1,35 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//====------ ConvertONNXToTorchPipeline.cpp - ONNX dialects to Torch lowering
+// pipeline -------===//
+//
+// Copyright 2019-2022 The IBM Research Authors.
+//
+// ================================================================================================
+//
+// This file registers the pipeline for converting ONNX to Torch Backend IR
+//
+//===-----------------------------------------------------------------------------------------===//
+
+#include "src/Conversion/ONNXToTorch/ONNXToTorchCommon.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+
+void registerONNXFrontendToTorchBackendPasses() {
+  PassPipelineRegistration<>("convert-onnx-to-torch-pipeline",
+      "Pipeline converting ONNX to Torch dialect.",
+      onnx_mlir::createONNXFrontendToTorchBackendPasses);
+}
+
+void createONNXFrontendToTorchBackendPasses(OpPassManager &pm) {
+  pm.addPass(createLowerToTorchPass());
+  pm.addPass(createFuncTorchTypeConversionPass());
+  pm.addPass(createFinalizingTorchTypeConversionPass());
+  pm.addPass(createEraseONNXEntryPointPass());
+}
+
+} // namespace onnx_mlir

src/Conversion/ONNXToTorch/EraseONNXEntryPoint.cpp

@@ -0,0 +1,59 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//====----- EraseModuleInitializer.cpp - ONNX dialects to Torch lowering
+//---------===//
+//
+// Copyright 2019-2022 The IBM Research Authors.
+//
+// ====================================================================================
+//
+// This file implements a pass for removing the ONNXEntryPointOp for
+// compatibility when converting to Torch backend IR.
+//
+//===------------------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/IR/BlockAndValueMapping.h"
+#include "mlir/IR/Builders.h"
+#include "mlir/IR/BuiltinOps.h"
+#include "mlir/Transforms/DialectConversion.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "src/Conversion/ONNXToTorch/ONNXToTorchCommon.hpp"
+
+using namespace mlir;
+using namespace onnx_mlir;
+
+namespace onnx_mlir {
+struct EraseONNXEntryPointPass
+    : public PassWrapper<EraseONNXEntryPointPass, OperationPass<ModuleOp>> {
+
+  StringRef getArgument() const override { return "erase-onnx-entry-point"; }
+
+  StringRef getDescription() const override {
+    return "Erase ONNXEntryPointOp.";
+  }
+
+  // Make sure that we have a valid default constructor and copy
+  // constructor to make sure that the options are initialized properly.
+  EraseONNXEntryPointPass() = default;
+  EraseONNXEntryPointPass(const EraseONNXEntryPointPass &pass)
+      : PassWrapper<EraseONNXEntryPointPass, OperationPass<ModuleOp>>() {}
+
+  void runOnOperation() override {
+    auto walkResult = getOperation().walk([](ONNXEntryPointOp op) {
+      op.erase();
+      return WalkResult::advance();
+    });
+    if (walkResult.wasInterrupted()) {
+      return signalPassFailure();
+    }
+  }
+};
+} // namespace onnx_mlir
+
+// std::unique_ptr<OperationPass<ModuleOp>>
+std::unique_ptr<mlir::Pass> onnx_mlir::createEraseONNXEntryPointPass() {
+  return std::make_unique<EraseONNXEntryPointPass>();
+}

src/Conversion/ONNXToTorch/Math/Elementwise.cpp

@@ -0,0 +1,49 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//===---------------- Elementwise.cpp - Elementwise Ops -------------------===//
+//
+// Copyright 2019-2022 The IBM Research Authors.
+//
+// =============================================================================
+//
+// This file lowers ONNX element-wise operators to Torch dialect.
+//
+//===----------------------------------------------------------------------===//
+
+#include "src/Conversion/ONNXToTorch/ONNXToTorchCommon.hpp"
+
+using namespace mlir;
+using namespace mlir::torch;
+
+namespace onnx_mlir {
+
+namespace {
+
+// AtenAddOp requires an additional alpha parameter and thus requires a unique
+// lowering
+class ConvertONNXAddOp : public OpConversionPattern<ONNXAddOp> {
+public:
+  using OpConversionPattern::OpConversionPattern;
+  LogicalResult matchAndRewrite(ONNXAddOp op, OpAdaptor adaptor,
+      ConversionPatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value one = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(1));
+    auto newResultType =
+        getTypeConverter()->convertType(op.getResult().getType());
+    rewriter.replaceOpWithNewOp<Torch::AtenAddTensorOp>(
+        op, newResultType, adaptor.A(), adaptor.B(), one);
+    return success();
+  }
+};
+} // namespace
+
+void populateLoweringONNXElementwiseOpToTorchPattern(
+    TypeConverter &typeConverter, RewritePatternSet &patterns,
+    MLIRContext *ctx) {
+  patterns.add<ConvertONNXAddOp>(typeConverter, ctx);
+}
+
+} // namespace onnx_mlir

src/Conversion/ONNXToTorch/ONNXToTorchCommon.cpp

@@ -0,0 +1,19 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+//====----- ONNXToTorchCommon.cpp - ONNX dialects to Torch lowering
+//---------===//
+//
+// Copyright 2019-2022 The IBM Research Authors.
+//
+// ===============================================================================
+//
+// This file contains common code shared by the functions performing the
+// lowering to the Torch backend dialect.
+//
+//===------------------------------------------------------------------------===//
+
+#include "src/Conversion/ONNXToTorch/ONNXToTorchCommon.hpp"
+
+namespace onnx_mlir {} // namespace onnx_mlir