compiler/plugins/input/Torch/PluginRegistration.cpp

// Copyright 2023 The IREE Authors
//
// Licensed under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception

#include "compiler/plugins/input/Torch/InputConversion/Passes.h"
#include "iree/compiler/Dialect/LinalgExt/IR/LinalgExtDialect.h"
#include "iree/compiler/PluginAPI/Client.h"
#include "mlir/Dialect/MLProgram/IR/MLProgram.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/Pass/PassManager.h"
#include "torch-mlir-dialects/Dialect/TMTensor/IR/TMTensorDialect.h"
#include "torch-mlir/Conversion/Passes.h"
#include "torch-mlir/Conversion/TorchOnnxToTorch/Passes.h"
#include "torch-mlir/Dialect/Torch/IR/TorchDialect.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionDialect.h"
#include "torch-mlir/Dialect/TorchConversion/Transforms/Passes.h"

namespace mlir::iree_compiler {

namespace {

struct TorchOptions {
  bool strictSymbolicShapes = true;
  bool decompose = true;
  void bindOptions(OptionsBinder &binder) {
    static llvm::cl::OptionCategory category("Torch Input");
    binder.opt<bool>(
        "iree-torch-use-strict-symbolic-shapes", strictSymbolicShapes,
        llvm::cl::cat(category),
        llvm::cl::desc("Forces dynamic shapes to be treated as strict"));
    binder.opt<bool>("iree-torch-decompose-complex-ops", decompose,
                     llvm::cl::cat(category),
                     llvm::cl::desc("Decompose complex torch operations."));
  }
};

// The torch plugin provides dialects, passes and opt-in options.
// Therefore, it is appropriate for default activation.
struct TorchSession
    : public PluginSession<TorchSession, TorchOptions,
                           PluginActivationPolicy::DefaultActivated> {
  static void registerPasses() {
    mlir::torch::registerTorchPasses();
    mlir::torch::registerTorchConversionPasses();
    mlir::torch::registerConversionPasses();
    mlir::torch::onnx_c::registerTorchOnnxToTorchPasses();
    TorchInput::registerTMTensorConversionPasses();
  }

  void onRegisterDialects(DialectRegistry &registry) override {
    registry.insert<torch::Torch::TorchDialect>();
    registry.insert<torch::TorchConversion::TorchConversionDialect>();
    registry.insert<mlir::torch::TMTensor::TMTensorDialect>();
    registry.insert<mlir::ml_program::MLProgramDialect>();
    registry.insert<IREE::LinalgExt::IREELinalgExtDialect>();
  }

  bool extendCustomInputConversionPassPipeline(
      OpPassManager &passManager, std::string_view typeMnemonic) override {
    if (typeMnemonic == "onnx") {
      // ONNX input is a pre-processing step to torch.
      mlir::torch::Torch::TorchLoweringPipelineOptions torchOnnxPipelineOptions;
      torchOnnxPipelineOptions.decompose = options.decompose;
      torchOnnxPipelineOptions.backendLegalOps =
          TorchInput::BackendLegalOps::get();
      mlir::torch::Torch::createTorchOnnxToTorchBackendPipeline(
          passManager, torchOnnxPipelineOptions);
    }

    if (typeMnemonic == "torch" || typeMnemonic == "onnx") {
      TorchInput::TorchToIREELoweringPipelineOptions torchOptions;
      torchOptions.strictSymbolicShapes = options.strictSymbolicShapes;
      torchOptions.decompose = options.decompose;
      TorchInput::createTorchToIREEPipeline(passManager, torchOptions);
      return true;
    }

    // TODO: Retire the tm_tensor input pipeline once we are fully switched
    // to the 'torch' pipeline, which handles everything from the 'torch'
    // dialect down (vs just 'tm_tensor' which was converting a couple of
    // ops to linalg).
    if (typeMnemonic == "tm_tensor") {
      passManager.addNestedPass<func::FuncOp>(
          TorchInput::createConvertTMTensorToLinalgExtPass());
      return true;
    }
    return false;
  }

  void populateCustomInputConversionTypes(StringSet<> &typeMnemonics) override {
    typeMnemonics.insert("tm_tensor");
    typeMnemonics.insert("torch");
    typeMnemonics.insert("onnx");
  }

  void populateDetectedCustomInputConversionTypes(
      ModuleOp &module, StringSet<> &typeMnemonics) override {
    auto *ctx = module.getContext();
    const Dialect *torchDialect = ctx->getLoadedDialect("torch");
    const Dialect *torchConversionDialect = ctx->getLoadedDialect("torch_c");
    const Dialect *tmTensorDialect = ctx->getLoadedDialect("tm_tensor");

    bool hasTorch = false;
    bool hasOnnx = false;
    // TODO: Retire the tm_tensor input pipeline
    bool hasTmTensor = false;

    module.walk([&](Operation *op) {
      Dialect *d = op->getDialect();
      if (d == torchDialect || d == torchConversionDialect) {
        hasTorch = true;
      } else if (d == tmTensorDialect) {
        hasTmTensor = true;
      }
      return WalkResult::advance();
    });

    for (auto funcOp : module.getOps<func::FuncOp>()) {
      if (funcOp->getAttrOfType<mlir::IntegerAttr>(
              "torch.onnx_meta.opset_version")) {
        hasOnnx = true;
        break;
      }
    }

    // ONNX is considered a superset of Torch. It runs all of the Torch
    // pipelines with an extra ONNX-specific preprocessing step.
    if (hasOnnx) {
      typeMnemonics.insert("onnx");
    } else if (hasTorch) {
      typeMnemonics.insert("torch");
    }

    if (hasTmTensor) {
      typeMnemonics.insert("tm_tensor");
    }
  }
};

} // namespace

} // namespace mlir::iree_compiler

IREE_DEFINE_COMPILER_OPTION_FLAGS(::mlir::iree_compiler::TorchOptions);

extern "C" bool iree_register_compiler_plugin_input_torch(
    mlir::iree_compiler::PluginRegistrar *registrar) {
  registrar->registerPlugin<::mlir::iree_compiler::TorchSession>("input_torch");
  return true;
}