torch-mlir/lib/Dialect/Torch/IR/TorchOps.cpp

//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, 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 "npcomp/Dialect/Torch/IR/TorchOps.h"

#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/PatternMatch.h"
#include "npcomp/Dialect/Basicpy/IR/BasicpyDialect.h"
#include "npcomp/Dialect/Basicpy/IR/BasicpyOps.h"
#include "npcomp/Dialect/Numpy/IR/NumpyDialect.h"
#include "npcomp/Dialect/Numpy/IR/NumpyOps.h"
#include "llvm/ADT/StringMap.h"

using namespace mlir;
using namespace mlir::NPCOMP;
using namespace mlir::NPCOMP::Torch;

static SmallVector<StringRef, 4> strArrayAttrToVector(ArrayAttr array) {
  SmallVector<StringRef, 4> strings;
  strings.reserve(array.size());
  for (auto stringAttr : array) {
    strings.push_back(stringAttr.cast<StringAttr>().getValue());
  }
  return strings;
}

//===----------------------------------------------------------------------===//
// KernelCallOp
//===----------------------------------------------------------------------===//

KernelMetadata KernelCallOp::getTorchKernelMetadata() {
  return KernelMetadata{
      .kernelName = kernelName(),
      .isVararg = sigIsVararg(),
      .isVarret = sigIsVarret(),
      .argTypes = strArrayAttrToVector(sigArgTypes()),
      .returnTypes = strArrayAttrToVector(sigRetTypes()),
  };
}

//===----------------------------------------------------------------------===//
// MethodOp
//===----------------------------------------------------------------------===//

LogicalResult MethodOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
  auto func = symbolTable.lookupNearestSymbolFrom<FuncOp>(*this, function());
  if (!func)
    return emitError() << "'@" << function()
                       << "' does not reference a valid function";
  if (func.getVisibility() != SymbolTable::Visibility::Private)
    return emitError() << "'@" << function()
                       << "' must reference a private function";
  if (func.isDeclaration())
    return emitError() << "'@" << function()
                       << "' must reference a function that is defined (not "
                          "merely declared)";
  auto expectedReceiverArgType = NnModuleType::get(
      getContext(), getOperation()->getParentOfType<ClassTypeOp>().getName());
  if (func.getType().getNumInputs() == 0 ||
      func.getType().getInput(0) != expectedReceiverArgType) {
    return emitError() << "the referenced function '" << function()
                       << "' must have a first argument of type "
                       << expectedReceiverArgType;
  }
  return success();
}

//===----------------------------------------------------------------------===//
// NnModuleOp
//===----------------------------------------------------------------------===//

static LogicalResult verify(NnModuleOp op) {
  for (Operation &child : *op.getBody())
    if (!isa<SlotOp, NnModuleTerminatorOp>(&child))
      return child.emitOpError() << "is not allowed inside 'torch.nn_module'";
  return success();
}

// PyTorch has a well-developed notion of subtyping.
//
// This is a restricted subset of it.
//
// TODO: Flesh this out.
bool isValidSubtype(Type subtype, Type type) {
  if (subtype == type)
    return true;
  if (auto optional = type.dyn_cast<OptionalType>())
    return subtype == optional.getContainedType() ||
           subtype.isa<Basicpy::NoneType>();
  return false;
}

LogicalResult NnModuleOp::verifySymbolUses(SymbolTableCollection &symbolTable) {
  auto classType =
      symbolTable.lookupNearestSymbolFrom<ClassTypeOp>(*this, getClassName());
  if (!classType)
    return emitError() << "'" << getClassName()
                       << "' does not reference a valid class type";

  auto attrs = llvm::to_vector<6>(getBody()->getOps<SlotOp>());
  auto attrDefs = llvm::to_vector<6>(classType.getBody()->getOps<AttrOp>());
  if (attrs.size() != attrDefs.size())
    return emitError() << "number of 'torch.slot's in a 'torch.nn_module' must "
                          "match number of 'torch.attr's in "
                          "the corresponding 'torch.class_type'";
  for (int i = 0, e = attrs.size(); i != e; i++) {
    SlotOp attr = attrs[i];
    AttrOp attrDef = attrDefs[i];
    if (!isValidSubtype(attr.value().getType(), attrDef.type()) ||
        attr.name() != attrDef.name()) {
      return attr.emitOpError()
          .append("is expected to match type and name of '",
                  attrDef.getOperation(), "'")
          .attachNote(attrDef.getLoc())
          .append("see torch.attr at corresponding index ", i, " here");
    }
  }

  return success();
}

//===----------------------------------------------------------------------===//
// ClassTypeOp
//===----------------------------------------------------------------------===//

static LogicalResult verify(ClassTypeOp op) {
  llvm::StringMap<Operation *> namesToOps;
  for (Operation &child : op.getBody()->without_terminator()) {
    if (!isa<AttrOp, MethodOp>(&child))
      return child.emitOpError() << "is not allowed inside `torch.class_type`";
    StringRef name;
    if (auto attr = dyn_cast<AttrOp>(child))
      name = attr.name();
    else
      name = cast<MethodOp>(child).name();
    auto itAndWasInserted = namesToOps.insert({name, &child});
    auto it = itAndWasInserted.first;
    bool wasInserted = itAndWasInserted.second;
    if (!wasInserted) {
      auto diag = op.emitOpError().append(
          "has duplicate attr/method with name '", name, "'");
      diag.attachNote(it->second->getLoc())
          .append("see first conflicting attr/method here");
      diag.attachNote(child.getLoc())
          .append("see second conflicting attr/method here");
      return failure();
    }
  }

  return success();
}

//===----------------------------------------------------------------------===//
// PrimLoopOp
//===----------------------------------------------------------------------===//

OperandRange PrimLoopOp::getSuccessorEntryOperands(unsigned index) {
  assert(index == 0);
  return iterArgsInit();
}

void PrimLoopOp::getSuccessorRegions(
    Optional<unsigned> index, ArrayRef<Attribute> operands,
    SmallVectorImpl<RegionSuccessor> &regions) {
  (void)operands;

  if (!index.hasValue()) {
    regions.emplace_back(&region(), region().getArguments().slice(1));
    return;
  }
  assert(*index == 0);
  regions.emplace_back(&region(), region().getArguments().slice(1));
  regions.emplace_back(getResults());
}

//===----------------------------------------------------------------------===//
// DerefineOp
//===----------------------------------------------------------------------===//

bool DerefineOp::areCastCompatible(mlir::TypeRange inputs,
                                   mlir::TypeRange outputs) {
  return isValidSubtype(inputs[0], outputs[0]);
}

void DerefineOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
                                             MLIRContext *context) {
  patterns.add(+[](DerefineOp op, PatternRewriter &rewriter) {
    // TODO: Properly model which ops allow type refinement.
    // For now, just assume all aten/torch ops allow refinement (which they do,
    // since that is what TorchScript IR allows).
    // See also: The comment in RefineTypes.cpp. For now, we copypasta from
    // there, since dependency-wise it's not clear where is the best place to
    // put this. Also, it seems like upstream MLIR might grow some useful
    // utilities to help with this case:
    // https://llvm.discourse.group/t/allow-shape-concretization-or-type-concretization-in-rewrites/3327/3
    // (or perhaps we should implement the AllowsTypeRefinement/Refinable
    // design in npcomp first and upstream it)
    //
    // TODO: Extend RefineTypes for this case and delete this canonicalization,
    // since we don't want control flow or calls to randomly block this fold
    // (this canonicalization pattern makes the compiler brittle to control flow
    // and calls).
    bool allAllowRefinement =
        llvm::all_of(op.getResult().getUsers(), [&](Operation *user) {
          StringRef ns = user->getDialect()->getNamespace();
          return ns == "aten" || ns == "torch";
        });
    if (!allAllowRefinement)
      return failure();
    rewriter.replaceOp(op, op.getOperand());
    return success();
  });
}


#define GET_OP_CLASSES
#include "npcomp/Dialect/Torch/IR/TorchOps.cpp.inc"