pytorch/torch/csrc/jit/codegen/cuda/fusion.cpp

#include <torch/csrc/jit/codegen/cuda/arith.h>
#include <torch/csrc/jit/codegen/cuda/codegen.h>
#include <torch/csrc/jit/codegen/cuda/fusion.h>
#include <torch/csrc/jit/codegen/cuda/fusion_segmenter.h>
#include <torch/csrc/jit/codegen/cuda/instrumentation.h>
#include <torch/csrc/jit/codegen/cuda/ir_all_nodes.h>
#include <torch/csrc/jit/codegen/cuda/ir_cloner.h>
#include <torch/csrc/jit/codegen/cuda/ir_printer.h>
#include <torch/csrc/jit/codegen/cuda/ir_utils.h>
#include <torch/csrc/jit/codegen/cuda/iter_visitor.h>
#include <torch/csrc/jit/codegen/cuda/kernel.h>
#include <torch/csrc/jit/codegen/cuda/lower2device.h>

namespace torch {
namespace jit {
namespace fuser {
namespace cuda {

static thread_local Fusion* ACTIVE_FUSION = nullptr; // NOLINT

FusionGuard::FusionGuard(Fusion* fusion) {
  prev_fusion = ACTIVE_FUSION;
  ACTIVE_FUSION = fusion;
}

FusionGuard::~FusionGuard() {
  ACTIVE_FUSION = prev_fusion;
}

Fusion* FusionGuard::getCurFusion() {
  return ACTIVE_FUSION;
}

void swap(Fusion& a, Fusion& b) noexcept {
  FUSER_PERF_SCOPE("Fusion swap");

  using std::swap;

  swap(static_cast<IrContainer&>(a), static_cast<IrContainer&>(b));

  swap(a.inputs_, b.inputs_);
  swap(a.outputs_, b.outputs_);

  swap(a.io_alias_, b.io_alias_);
  swap(a.permuted_input_map_, b.permuted_input_map_);
  swap(a.permuted_output_map_, b.permuted_output_map_);
}

std::unique_ptr<SegmentedFusion> Fusion::segment(
    const at::ArrayRef<IValue>& inputs) {
  FUSER_PERF_SCOPE("Segment Fusion");
  return SegmentCandidateFinder::segment(this, inputs);
}

IrCloner Fusion::copy(const Fusion* from, Fusion* to) {
  to->clear();
  auto ir_cloner = IrContainer::copy(from, to);

  for (auto val : from->vals_) {
    ir_cloner.clone(val)->setDefinition(ir_cloner.clone(val->definition_));
    ir_cloner.clone(val)->setUses(ir_cloner.clone(val->uses_));
  }

  to->inputs_ = ir_cloner.clone(from->inputs_);
  to->outputs_ = ir_cloner.clone(from->outputs_);

  // TODO: put this into ir_cloner instead
  for (const auto& entry : from->io_alias_) {
    Val* copied_output = ir_cloner.clone(entry.first);
    Val* copied_input = ir_cloner.clone(entry.second);
    to->io_alias_[copied_output] = copied_input;
  }

  to->permuted_input_map_ = from->permuted_input_map_;
  to->permuted_output_map_ = from->permuted_output_map_;

  to->all_tv_uses_valid_ = from->all_tv_uses_valid_;
  // This should never be true on copy, but copying for completeness.
  to->is_during_update_uses_ = from->is_during_update_uses_;

  return ir_cloner;
}

// Clang tidy complains when using default constructor for IrContainer instead
// of copy constructor. Fusion::copy has a call to IrContainer::copy, so it's
// redundant to use the IrContainer copy constructor, but it is harmless since
// Fusion::copy starts by calling clear().
Fusion::Fusion(const Fusion& other) : IrContainer(other) {
  FUSER_PERF_SCOPE("Fusion copy");
  Fusion::copy(&other, this);
}

Fusion::Fusion(Fusion&& other) noexcept {
  FUSER_PERF_SCOPE("Fusion move");
  swap(*this, other);
}

Fusion& Fusion::operator=(const Fusion& other) {
  FUSER_PERF_SCOPE("Fusion copy assign");
  Fusion copy(other);
  clear();
  swap(*this, copy);
  return *this;
}

Fusion& Fusion::operator=(Fusion&& other) noexcept {
  FUSER_PERF_SCOPE("Fusion move assign");
  clear();
  swap(*this, other);
  return *this;
}

Fusion::~Fusion() {
  clear();
}

void Fusion::clear() noexcept {
  FUSER_PERF_SCOPE("Fusion clear");

  IrContainer::clear();

  inputs_.clear();
  outputs_.clear();

  io_alias_.clear();

  permuted_input_map_.clear();
  permuted_output_map_.clear();

  all_tv_uses_valid_ = false;
  is_during_update_uses_ = false;
}

void Fusion::removeExpr(Expr* expr) {
  assertInContainer(expr, "Cannot remove expr ");
  // If we hit this error too frequently, we could lighten the restrictions so
  // that removing something that doesn't exist simply does nothing. For now,
  // we're going with the strictest model which errors.

  for (auto out : expr->outputs()) {
    out->setDefinition(nullptr);
  }

  for (auto inp : expr->inputs()) {
    auto uses_copy = inp->uses();
    auto it = std::find(uses_copy.begin(), uses_copy.end(), expr);
    if (it != uses_copy.end()) {
      uses_copy.erase(it);
      inp->setUses(uses_copy);
    }
  }

  IrContainer::removeExpr(expr);
}

void Fusion::removeVal(Val* val) {
  assertInContainer(val, "Cannot remove val ");

  TORCH_CHECK(
      !val->isFusionInput(),
      "Cannot remove val as it is an input of the fusion.");
  TORCH_CHECK(
      !val->isFusionOutput(),
      "Cannot remove val as it is an output of the fusion.");

  Expr* orig = val->definition();
  if (orig != nullptr)
    removeExpr(val->definition());

  for (Expr* use : unordered_uses(val)) {
    removeExpr(use);
  }
  IrContainer::removeVal(val);
}

void Fusion::addInput(Val* input) {
  assertInContainer(input, "Cannot register input ");

  TORCH_INTERNAL_ASSERT(
      input->getDataType() != DataType::Index,
      "Data type Index is a local compile time data type only, it cannot be used as an input in case it was generated from another kernel.");

  if (input->getValType().value() == ValType::TensorView) {
    auto tv = input->as<TensorView>();
    tv->setMemoryType(MemoryType::Global);
  } else if (input->getValType().value() == ValType::Scalar) {
    TORCH_CHECK(
        !input->isConst(),
        "Immediate scalar value cannot be added as an input. It is not necessary to pass it as an input.");
  }

  inputs_.push_back(input);
  input->setIsFusionInput(true);

  all_tv_uses_valid_ = false;
}

void Fusion::addOutput(Val* output) {
  // We currently don't support explicitly outputing aliased inputs. This is
  // because they are already marked as output for in-place update. It's tricky
  // to allow marking them explicitly as real output, since that requires us to
  // register/identify output not only by `Val*` pointer, but also by indices;
  // it also requires us to magically arrange `outputs_` entries in proper order
  // ^^^ this doesn't look intuitive on `outputs_` in fusion.
  // I think we can solve this by marking addOutput on io_alias_ keys after
  // fusion is fully defined. Tracking this in #1488
  // Apparently we can't do this neither at the time. I think segmentation
  // unfortunately would call addOutput after we marked io_alias_ map.
  // TORCH_CHECK(io_alias_.count(output) == 0,
  //     "can't register aliased output as real output");

  assertInContainer(output, "Cannot register output ");
  if (output->getValType().value() == ValType::TensorView) {
    auto tv = output->as<TensorView>();
    tv->setMemoryType(MemoryType::Global);
  }
  outputs_.push_back(output);
  output->setIsFusionOutput(true);

  all_tv_uses_valid_ = false;
}

void Fusion::addOutput(WelfordResult& wr) {
  // Want to always make sure the avg gets added last
  //  since avg will be the out() value of welfordOp,
  //  and want to make it the top of the computeAt chain
  addOutput(wr.var_sum);
  addOutput(wr.n);
  addOutput(wr.avg);
}

void Fusion::removeInput(Val* input) {
  auto find_input = std::find(inputs_.begin(), inputs_.end(), input);
  if (find_input != inputs_.end()) {
    inputs_.erase(find_input);
  }
  input->setIsFusionInput(false);
  all_tv_uses_valid_ = false;
}

void Fusion::removeOutput(Val* output) {
  auto find_output = std::find(outputs_.begin(), outputs_.end(), output);
  if (find_output != outputs_.end()) {
    outputs_.erase(find_output);
  }
  output->setIsFusionOutput(false);
  all_tv_uses_valid_ = false;
}

void Fusion::replaceOutput(Val* output, Val* replacement) {
  auto find_output = std::find(outputs_.begin(), outputs_.end(), output);
  TORCH_CHECK(find_output != outputs_.end(), "Unable to find output in Fusion");

  if (find_output != outputs_.end()) {
    *find_output = replacement;

    if (replacement->getValType().value() == ValType::TensorView) {
      replacement->setIsFusionOutput(true);
      replacement->as<TensorView>()->setMemoryType(MemoryType::Global);
    }
    if (output->getValType().value() == ValType::TensorView) {
      output->setIsFusionOutput(false);
      output->as<TensorView>()->setMemoryType(MemoryType::Local);
    }
    resetTvUses();
  }

  // Temporary WAR for issue #1112
  // (https://github.com/csarofeen/pytorch/issues/1112)
  if (io_alias_.count(output) != 0) {
    auto input = io_alias_[output];
    io_alias_.erase(output);
    io_alias_[replacement] = input;
  }
}

std::vector<Expr*> Fusion::exprs() {
  return StmtSort::getExprs(this);
}

std::vector<Val*> Fusion::inputsOf(Val* val) {
  return InputsOf::output(this, val);
}

void Fusion::validateInputs() {
  std::unordered_set<Val*> all_inputs;
  for (Val* out : outputs()) {
    for (Val* input : inputsOf(out)) {
      all_inputs.insert(input);
    }
  }

  std::unordered_set<Val*> input_dims;
  auto inp_tvs = ir_utils::filterByType<TensorView>(inputs());
  for (auto tv : inp_tvs) {
    for (auto id : tv->getMaybeRFactorDomain()) {
      input_dims.emplace(id->extent());
    }
  }
  for (Val* input : all_inputs) {
    if (!input->isConstScalar()) {
      TORCH_CHECK(
          input->isFusionInput() ||
              // TODO: Switch:
              inContainer(input),
          // to: input_dims.find(input) != input_dims.end(),
          // https://github.com/csarofeen/pytorch/issues/1365
          "Could not figure out how ",
          input->toString(),
          " is generated, however it was not specified as an input.");
    }
  }
}

void Fusion::print() {
  FUSER_PERF_SCOPE("Fusion::print");

  FusionGuard fg(this);
  std::cout << "\n%kernel {\n";
  IrMathPrinter op_exprs(std::cout);
  op_exprs.handle(this);
  std::cout << "\nTransformPrinter : \n";
  IrTransformPrinter t_exprs(std::cout);
  t_exprs.handle(this);
  std::cout << "}\n\n";
}

void Fusion::printKernel(DataType index_type) {
  FUSER_PERF_SCOPE("Fusion::printKernel");
  TORCH_INTERNAL_ASSERT(
      !this->isA<kir::Kernel>(),
      "Cannot \"print kernel\" of a kernel container. ",
      "This would require lowering during lowering.");
  std::cout << codegen::generateCudaKernel(GpuLower(this, index_type).kernel());
}

void Fusion::printMath(bool from_outputs_only) {
  FUSER_PERF_SCOPE("Fusion::printMath");

  FusionGuard fg(this);
  auto exprs_for_print = exprs();
  std::cout << "Inputs:" << std::endl;
  for (auto inp : inputs()) {
    std::cout << "  " << inp << ", " << inp->getDataType().value() << std::endl;
  }

  std::cout << "Outputs:" << std::endl;
  for (auto out : outputs()) {
    std::cout << "  " << out << ", " << out->getDataType().value() << std::endl;
  }

  // If we want everything in the fusion, grab all values without uses to
  // traverse from.
  if (!from_outputs_only) {
    std::vector<Val*> leaf_vals;
    for (auto val : deterministic_vals()) {
      if (val->uses().empty()) {
        leaf_vals.push_back(val);
      }
    }
    exprs_for_print = StmtSort::getExprs(this, leaf_vals);
  }

  std::cout << "\n%kernel_math {\n";
  for (auto expr : exprs_for_print) {
    std::cout << expr;
  }
  std::cout << "}\n\n";
}

void Fusion::printTransforms() {
  FUSER_PERF_SCOPE("Fusion::printTransforms");

  FusionGuard fg(this);
  IrTransformPrinter t_exprs(std::cout);
  t_exprs.handle(this);
}

void Fusion::registerVal(Val* val) {
  if (inContainer(val)) {
    return;
  }

  if (val->fusion()) {
    TORCH_CHECK(
        val->fusion() == this, val, " was not found in the active fusion.");
  }

  IrContainer::registerVal(val);
}

void Fusion::registerExpr(Expr* expr) {
  if (inContainer(expr)) {
    return;
  }

  if (expr->fusion()) {
    TORCH_CHECK(
        expr->fusion() == this, expr, " was not found in the active fusion.");
  }

  IrContainer::registerExpr(expr);

  bool has_tv = false;

  for (Val* input : expr->inputs()) {
    has_tv = has_tv || input->isA<TensorView>();
    assertInContainer(input, "Input to expr is invalid, ");
    auto uses_copy = input->uses();
    if (std::find(uses_copy.begin(), uses_copy.end(), expr) ==
        uses_copy.end()) {
      uses_copy.push_back(expr);
      input->setUses(uses_copy);
    }
  }

  // Kernel is the only container type that is non-ssa. This is mainly (maybe
  // only) because of initialization expressions which would overwrite tensor
  // view definitions.
  bool is_ssa = !this->isA<kir::Kernel>();

  for (Val* output : expr->outputs()) {
    has_tv = has_tv || output->isA<TensorView>();
    assertInContainer(output, "Output to expr is invalid, ");
    if (output->definition() != nullptr && is_ssa) {
      removeExpr(output->definition());
    }
    if (is_ssa || (!is_ssa && output->definition() == nullptr)) {
      output->setDefinition(expr);
    }
  }

  if (has_tv) {
    resetTvUses();
  }
}

void Fusion::resetTvUses() {
  FUSER_PERF_SCOPE("Fusion::resetTvUses");
  is_during_update_uses_ = true;

  // getExprs only uses definition, so even if we've modified uses already to
  // remove dead exprs, this could reinsert them. getExprs is also boundeds by
  // inputs as registered inputs will return nullptr as their definition.
  const auto all_tvs = ir_utils::filterByType<TensorView>(vals_);
  const auto used_exprs = StmtSort::getExprs(this);

  for (auto tv : all_tvs) {
    tv->setUses({});
  }

  // Same as in register expr
  for (auto expr : used_exprs) {
    for (Val* input : expr->inputs()) {
      auto uses_copy = input->uses();
      if (std::find(uses_copy.begin(), uses_copy.end(), expr) ==
          uses_copy.end()) {
        uses_copy.push_back(expr);
        input->setUses(uses_copy);
      }
    }
  }

  all_tv_uses_valid_ = true;
  is_during_update_uses_ = false;
}

std::vector<Val*> Fusion::usedMathVals() {
  // Note that using fusion->inputs() as the argument for the first
  // parameter of getAllValsBetween does not grab all used vals as
  // there can be vals that are created inside a fusion without using
  // anything from inputs. See, for example, tv0 in the
  // FusionOuterSplit test.
  const auto inputs = InputsOf::outputs(this, outputs());
  auto used_math_vals = DependencyCheck::getAllValsBetween(
      {inputs.begin(), inputs.end()}, outputs());
  // When an expre has multiple outputs and only some of them are
  // used, the rest aren't included in used_math_vals as they are not
  // used. However, we want them to be included as they must show up
  // in the fusion.
  std::vector<Val*> vals_to_add;
  std::unordered_set<Val*> added_vals;

  for (auto val : used_math_vals) {
    auto def = val->definition();
    if (def == nullptr || def->outputs().size() < 2) {
      continue;
    }
    for (auto out : def->outputs()) {
      if (std::find(used_math_vals.begin(), used_math_vals.end(), out) ==
          used_math_vals.end()) {
        if (!added_vals.count(out)) {
          vals_to_add.push_back(out);
          added_vals.insert(out);
        }
      }
    }
  }

  used_math_vals.insert(
      used_math_vals.end(), vals_to_add.begin(), vals_to_add.end());

  return used_math_vals;
}

std::unordered_set<Expr*> Fusion::unordered_uses(Val* val) const {
  return std::unordered_set<Expr*>(val->uses().begin(), val->uses().end());
}

Expr* Fusion::definition(const Val* val) const {
  assertInContainer(val, "Cannot detect the definition of val, ");
  return val->definition();
}

// Indicate to kernel to set itself up to generate random numbers
bool Fusion::isStochastic() {
  for (auto expr : exprs())
    if (expr->getExprType() == ExprType::UnaryOp)
      if (expr->as<UnaryOp>()->getUnaryOpType() == UnaryOpType::RandLike)
        return true;
  return false;
}

std::vector<Val*> Fusion::getTerminatingOutputs() {
  FUSER_PERF_SCOPE("getTerminatingOutputs");

  auto is_reachable_to_output = [](Val* val) {
    // traverse to consumers of val and see if there is an output
    std::deque<Val*> consumers;
    for (auto use : val->uses()) {
      for (auto consumer : use->outputs()) {
        consumers.push_back(consumer);
      }
    }
    while (!consumers.empty()) {
      auto consumer = consumers.back();
      consumers.pop_back();
      if (consumer->isFusionOutput()) {
        return true;
      }
      // consumer is not an output; proceed to its consumers
      for (auto use : consumer->uses()) {
        for (auto consumer_of_consumer : use->outputs()) {
          consumers.push_back(consumer_of_consumer);
        }
      }
    }
    return false;
  };

  std::vector<Val*> terminating_outputs;

  for (auto out : outputs()) {
    // If there is another output reachable from this output, it's not
    // terminating.
    if (is_reachable_to_output(out)) {
      continue;
    }
    terminating_outputs.push_back(out);
  }

  return terminating_outputs;
}

bool Fusion::isAliasCompatible(Val* left, Val* right) {
  // Nullptr check
  if (left == nullptr || right == nullptr) {
    return false;
  }

  // DataType check
  if (!left->getDataType().has_value() || !right->getDataType().has_value() ||
      left->getDataType().value() != right->getDataType().value()) {
    return false;
  }

  // ValType check
  if (!left->getValType().has_value() || !right->getValType().has_value() ||
      left->getValType().value() != right->getValType().value()) {
    return false;
  }

  // Check same number of dimensions if both values are TensorViews
  if (ir_utils::isTV(left) && ir_utils::isTV(right)) {
    return left->as<TensorView>()->nDims() == right->as<TensorView>()->nDims();
  }
  return false;
}

void Fusion::aliasOutputToInput(Val* output, Val* input) {
  // Because we could cast output when input is casted.
  TORCH_INTERNAL_ASSERT(
      !output->isFusionOutput(),
      "Do NOT add aliased output to fusion output outside of `aliasOutputToInput");

  if (!input->isFusionInput()) {
    auto input_expr = input->definition();
    // TORCH_INTERNAL_ASSERT(input_def.etype() == ExprType::UnaryOp, "expected
    // unary op for aliased input");
    TORCH_INTERNAL_ASSERT(
        input_expr->isA<UnaryOp>(), "expected unary op for aliased input");
    auto input_uop = input_expr->as<UnaryOp>();
    TORCH_INTERNAL_ASSERT(
        input_uop->getUnaryOpType() == UnaryOpType::Cast,
        "expected aliased input to be output of cast op");
    input = input_uop->in();
  }
  TORCH_INTERNAL_ASSERT(
      input->getDataType().has_value() && output->getDataType().has_value(),
      "requires DataType to be available for aliased output to input");

  if (input->getDataType().value() != output->getDataType().value()) {
    output = castOp(input->getDataType().value(), output);
  }
  // TODO: output should be marked at the end of fusion definition #1488
  addOutput(output);

  TORCH_INTERNAL_ASSERT(
      isAliasCompatible(input, output),
      "The input and output values are not alias-compatible.");
  io_alias_[output] = input;
}

Val* Fusion::getOutputAlias(Val* output) {
  auto search = io_alias_.find(output);
  if (search != io_alias_.end()) {
    return search->second;
  }
  return nullptr;
}

std::unordered_set<int> Fusion::getOutputAliasIndices() const {
  if (io_alias_.empty()) {
    return {};
  }

  std::unordered_set<int> alias_indices;

  for (const auto i : c10::irange(outputs_.size())) {
    if (io_alias_.count(outputs_[i]) != 0) {
      alias_indices.insert(i);
    }
  }
  return alias_indices;
}

std::vector<std::pair<int, int>> Fusion::getInputAliasIndices() const {
  if (io_alias_.empty()) {
    return {};
  }

  std::vector<std::pair<int, int>> alias_indices;
  for (const auto i : c10::irange(outputs_.size())) {
    if (io_alias_.count(outputs_[i]) != 0) {
      bool found = false;
      for (const auto j : c10::irange(inputs_.size())) {
        if (io_alias_.at(outputs_[i]) == inputs_[j]) {
          alias_indices.emplace_back(i, j);
          found = true;
          break;
        }
      }
      TORCH_INTERNAL_ASSERT(
          found,
          "io_alias_ mapping failure, alias output is not present in inputs");
    }
  }
  // can't assert here, we could have segmented fusion where not all alias
  // outputs are present

  return alias_indices;
}

} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch