pytorch/torch/csrc/jit/passes/symbolic_shape_analysis.cpp

#include <ATen/core/symbol.h>
#include <c10/util/Exception.h>
#include <c10/util/irange.h>
#include <torch/csrc/jit/ir/alias_analysis.h>
#include <torch/csrc/jit/ir/constants.h>
#include <torch/csrc/jit/ir/ir.h>
#include <torch/csrc/jit/ir/ir_views.h>
#include <torch/csrc/jit/jit_log.h>
#include <torch/csrc/jit/passes/common_subexpression_elimination.h>
#include <torch/csrc/jit/passes/constant_pooling.h>
#include <torch/csrc/jit/passes/constant_propagation.h>
#include <torch/csrc/jit/passes/dead_code_elimination.h>
#include <torch/csrc/jit/passes/integer_value_refinement.h>
#include <torch/csrc/jit/passes/loop_unrolling.h>
#include <torch/csrc/jit/passes/lower_tuples.h>
#include <torch/csrc/jit/passes/peephole.h>
#include <torch/csrc/jit/passes/peephole_list_idioms.h>
#include <torch/csrc/jit/passes/peephole_non_tensor.h>
#include <torch/csrc/jit/passes/remove_mutation.h>
#include <torch/csrc/jit/passes/shape_analysis.h>
#include <torch/csrc/jit/passes/symbolic_shape_analysis.h>
#include <torch/csrc/jit/passes/symbolic_shape_cache.h>
#include <torch/csrc/jit/passes/tensorexpr_fuser.h>
#include <torch/csrc/jit/runtime/exception_message.h>
#include <torch/csrc/jit/runtime/symbolic_shape_registry.h>
#include <algorithm>
#include <memory>
#include <numeric>
#include <unordered_map>
#include <utility>
#include <vector>

/*
XXX: this is still in prototype phase and has much work left to do, including
but not limited to:
- Refactor APIs
- Add decent coverage of common ops
- Add shape analysis pass on Graph that handles Loops
- Allow concurrent reads to the operator map
- Supporting returning partially evaluated shape compute graph
*/

static bool symbolic_shape_analysis_test_mode = false;

namespace torch {
namespace jit {

// This is similar to c10::SymbolicShape, but instead of either having
// a concrete dimension or a symbolic dimension, an argument may be:
// - A Symbolic Dimension
// - A Constant Integer
// - Neither of the above. The third case can occur due to inputs to
// ops like view that accept negative values. Maintaining the distinction
// between an unknown symbolic dimension and an unknown integer allows
// us to optimize out comparisons to values < 0 (symbolic shapes are always >=
// 0) For example, a call like graph(%y: Tensor(SS(-1), 10, 10), %inp: int):
//   %five: int = prim::Constant[value=5]()
//   %zero: int = prim::Constant[value=0]()
//   %1 : int = aten::size(%y, %zero)
//   %2 : int[] = prim::ListConstruct(%five, %1, %inp)
//   %y.2: Tensor(5, SS(-1), (New Symbolic Shape)) = aten::view(%y, %2)
//
// x.view([5, y.size(0), inp])
// will have inputs equal to [5, SS(-1), std::nullopt]

struct ShapeArg
    : public std::
          pair<std::optional<c10::ShapeSymbol>, std::optional<int64_t>> {
  using pair::pair;

  static ShapeArg unknownInteger() {
    return ShapeArg();
  }

  ShapeArg(int64_t int_value) {
    this->first = std::nullopt;
    this->second = int_value;
  }

  ShapeArg(c10::ShapeSymbol ss) {
    if (ss.is_static()) {
      this->first = std::nullopt;
      this->second = ss.value();
    } else {
      this->first = ss;
      this->second = std::nullopt;
    }
  }

  std::optional<int64_t> asConstantInt() const {
    return this->second;
  }

  std::optional<c10::ShapeSymbol> asShapeSymbol() const {
    return this->first;
  }

 private:
  ShapeArg() {
    this->first = std::nullopt;
    this->second = std::nullopt;
  }
};

static std::ostream& operator<<(std::ostream& out, const ShapeArg& sa) {
  if (auto val = sa.asConstantInt()) {
    out << *val;
  } else if (auto ss = sa.asShapeSymbol()) {
    out << *ss;
  } else {
    out << "UNK";
  }
  return out;
}

struct ShapeArguments {
  // Superset of SymbolicShape, with additional support for unknown, nonsymbolic
  // vals
 public:
  ShapeArguments(const c10::SymbolicShape& ss) {
    has_dim_ = ss.rank().has_value();
    if (has_dim_) {
      for (size_t i = 0; i < *ss.rank(); ++i) {
        maybe_shape_symbols_.emplace_back(ss.at(i));
      }
    }
  }

  ShapeArguments(std::vector<ShapeArg> ss)
      : has_dim_(true), maybe_shape_symbols_(std::move(ss)) {}

  bool has_dim() const {
    return has_dim_;
  }

  int64_t len() const {
    TORCH_INTERNAL_ASSERT(has_dim_, "ShapeArguments has no known dim")
    return (int64_t)maybe_shape_symbols_.size();
  }

  const ShapeArg at(size_t i) const {
    TORCH_INTERNAL_ASSERT(has_dim_, "ShapeArguments has no known dim")
    return maybe_shape_symbols_.at(i);
  }

 private:
  bool has_dim_;
  std::vector<ShapeArg> maybe_shape_symbols_;
};

static std::ostream& operator<<(std::ostream& os, const ShapeArguments& sa) {
  if (!sa.has_dim()) {
    os << "(UNKNOWN DIM)";
    return os;
  }

  os << "(";
  for (const auto i : c10::irange(sa.len())) {
    os << sa.at(i);
  }
  os << ")";

  return os;
}

bool setSymbolicShapeAnalysisTestMode(bool value) {
  bool old_value = symbolic_shape_analysis_test_mode;
  symbolic_shape_analysis_test_mode = value;
  return old_value;
}

bool symbolicShapeAnalysisTestModeEnabled() {
  return symbolic_shape_analysis_test_mode;
}

using SSArgument = std::variant<ShapeArguments, IValue>;

static std::ostream& operator<<(std::ostream& out, const SSArgument& sa) {
  if (const IValue* iv = std::get_if<IValue>(&sa)) {
    out << *iv;
  } else {
    out << std::get<ShapeArguments>(sa);
  }
  return out;
}

namespace {

bool isListOfInts(const TypePtr& type) {
  return type->cast<ListType>() &&
      type->cast<ListType>()->getElementType()->cast<IntType>();
}

bool isListOfListOfInts(const TypePtr& type) {
  // Allows List[Optional[List[Int]]]
  if (!type->cast<ListType>()) {
    return false;
  }
  TypePtr element_type = type->cast<ListType>()->getElementType();
  if (element_type->cast<OptionalType>()) {
    element_type = element_type->cast<OptionalType>()->getElementType();
  }
  return isListOfInts(element_type);
}

bool isListOfTensors(const TypePtr& type) {
  return type->cast<ListType>() &&
      type->cast<ListType>()->getElementType()->cast<TensorType>();
}

std::optional<size_t> normIndex(int64_t index, size_t len) {
  if (index < 0) {
    index = index + len;
  }
  if (index >= 0 && index < static_cast<int64_t>(len)) {
    return index;
  } else {
    return std::nullopt;
  }
}

bool shapeGraphCleanupPasses(std::shared_ptr<Graph> graph) {
  // TODO: lower simple tuples ?
  bool made_change = RemoveListMutation(graph);
  made_change |= UnrollConstantLoops(graph);
  made_change |= ConstantPropagation(graph);
  made_change |= PeepholeOptimizeNonTensor(graph);
  made_change |= PeepholeOptimizeListIdioms(graph, /*refine_list_len*/ true);
  made_change |= RefineIntegerValues(graph);
  made_change |= ConstantPropagation(graph);
  // todo add return change for constant pooling
  ConstantPooling(graph);
  made_change |= EliminateCommonSubexpression(graph);
  EliminateDeadCode(graph);
  return made_change;
}

void replaceWithIValue(Value* v, IValue val) {
  WithInsertPoint guard(*v->node()->owningBlock()->nodes().begin());
  v->replaceAllUsesWith(v->owningGraph()->insertConstant(val));
}

c10::SymbolicShape extractListShape(
    Value* list,
    std::unordered_map<Value*, int64_t>& symbolic_shape_values,
    const AliasDb& db) {
  if (list->node()->kind() == prim::Constant) {
    auto int_list = toIValue(list)->toIntVector();
    return c10::SymbolicShape(int_list);
  }
  // We need a list construct or a constant output
  // that is not written to in order to analyze the output shape
  if (list->node()->kind() != prim::ListConstruct || db.hasWriters(list)) {
    GRAPH_DEBUG("Could not extract shape");
    return c10::SymbolicShape();
  }
  Node* list_construct = list->node();
  std::vector<std::optional<int64_t>> output_shape;
  for (Value* input : list_construct->inputs()) {
    if (symbolic_shape_values.count(input)) {
      output_shape.emplace_back(symbolic_shape_values[input]);
    } else {
      output_shape.push_back(constant_as<int64_t>(input));
    }
  }
  return c10::SymbolicShape(output_shape);
}

// Symbolic Shape Analysis works through iteratively partially evaluating
// a TorchScript shape compute graph by inputting properties from input
// Tensors. We can substitute in properties like `len(x)` and `x[1]`
// if they are statically on the input Tensors. We can also use
// assertions like `assert len(x) == 4` in order to refine the input
// length and unroll loops over its elements. We iteratively optimize and
// substitute in properties until we are unable to make any further
// optimizations. Finally, we try to extract Tensor properties from the output.
// For instance `return [1, 2, inp[2] + 1, inp[3]]` we know that the output
// will be length 4 with first two dimensions equal to 1 and 2. We can also
// deduce that the 4th dimension has the same symbolic shape as inp[3], which
// means that we do know its concrete value statically but we can assign sets
// of tensor dimensions which must be equal at runtime.

struct SymbolicShapeOpAnalyzer {
  std::shared_ptr<Graph> shape_compute_graph_;
  const FunctionSchema* schema_;
  std::vector<SSArgument> inputs_;

  // For the case where we have a JIT graph,
  // substitute optional types for their component types
  // if the type is known. This doesn't need to be done
  // for known IValues.
  void refineInputUnionTypes(const Node* parent_graph_node) {
    for (size_t op_in_index = 0;
         op_in_index < shape_compute_graph_->inputs().size();
         op_in_index++) {
      auto type = parent_graph_node->input(op_in_index)->type();
      if (auto opt_type = shape_compute_graph_->inputs()
                              .at(op_in_index)
                              ->type()
                              ->cast<OptionalType>()) {
        // None will get handled with constant substitution later
        if (!type->cast<OptionalType>() &&
            !NoneType::get()->isSubtypeOf(*type)) {
          shape_compute_graph_->inputs()
              .at(op_in_index)
              ->setType(opt_type->getElementType());
        }
      } else if (shape_compute_graph_->inputs()
                     .at(op_in_index)
                     ->type()
                     ->cast<NumberType>()) {
        shape_compute_graph_->inputs().at(op_in_index)->setType(type);
      }
    }
  }

  // We handle non-constant values in the shape propagation step
  void substituteConstantInputs() {
    if (shape_compute_graph_->inputs().empty()) {
      return;
    }

    bool seen_tensor_list = false;

    size_t op_in_index = 0;
    while (op_in_index < shape_compute_graph_->inputs().size()) {
      Value* graph_in_var = shape_compute_graph_->inputs().at(op_in_index);
      if (!isListOfListOfInts(graph_in_var->type())) {
        op_in_index++;
        continue;
      }

      // Modifying the graph where _node is part of to not use the tensor
      // construct

      // When we have partially evaluate a list of Tensors like cat(tensor[])
      // We have a few problems:
      // - optimizing out calls to the length of the list: len(tensors)
      // - resolving accesses of the list to the tensor symbolic sizes the
      // corresponding list element We can solve both of these problems by
      // replacing the partial evaluation of cat([x, y]) def cat(tensors:
      // List[List[int]], dim: int)
      //    body
      // with
      // def cat(x, y, dim: int)
      //     tensors = [x, y]
      //     body
      TORCH_INTERNAL_ASSERT(
          !seen_tensor_list,
          "SSA doesn't handle case with multiple tensor lists")
      seen_tensor_list = true;

      uint64_t li_length = inputs_.size() - (schema_->arguments().size() - 1);
      std::vector<Value*> li_inputs;

      TypePtr element_type =
          graph_in_var->type()->cast<ListType>()->getElementType();
      for (size_t j = op_in_index; j < op_in_index + li_length; ++j) {
        auto new_inp = shape_compute_graph_->insertInput(op_in_index + j);
        new_inp->setType(element_type);
        li_inputs.push_back(new_inp);
      }
      WithInsertPoint guard(*shape_compute_graph_->block()->nodes().begin());
      auto new_li = shape_compute_graph_->insertNode(
          shape_compute_graph_->createList(element_type, li_inputs));
      graph_in_var->replaceAllUsesWith(new_li->output());
      shape_compute_graph_->eraseInput(op_in_index + li_length);
    }

    TORCH_INTERNAL_ASSERT(
        shape_compute_graph_->inputs().size() <= inputs_.size(),
        "Shape Compute Graph expected to have less inputs than actual inputs"); //?

    for (size_t op_in_index = 0;
         op_in_index < shape_compute_graph_->inputs().size();
         op_in_index++) {
      SSArgument& argument = inputs_[op_in_index];
      Value* graph_in_var = shape_compute_graph_->inputs().at(op_in_index);

      if (IValue* cur_val = std::get_if<IValue>(&argument)) {
        GRAPH_DEBUG("Substituting constant input ", *cur_val);
        replaceWithIValue(graph_in_var, *cur_val);
      } else {
        auto cur_arg = std::get<ShapeArguments>(argument);
        if (cur_arg.has_dim()) {
          graph_in_var->setType(ListType::ofInts());
        }
      }
    }
  }

  void substituteSymbolicProperties(
      std::unordered_map<Value*, int64_t>* symbolic_shape_values) {
    // clang-format off
    // here we iteratively substitute properties of the node's input tensors
    // into the shape compute graph. we can substitute constants into the
    // like len(inp) or inp[0] if the tensor has a fixed length or a fixed
    // first dimension. we also try to resolve symbolic shapes of the same
    // symbolic value to the same Value * in the shape compute graph.
    // for the shape logic:
    // dim1 = inp1[0]
    // dim2 = inp2[0]
    // return dim1 if dim2 == 1 else dim2
    // if we see that inp1[0] and inp2[0] both have the same symbolic shape
    // value, then it is a valid transformation to replace dim2 with dim1 or
    // vice versa. to do this we collect all Value * for a particular symbolic
    // shape. Then, we replace all Value * within that set with their dominator.
    // In the example above, this allows us to infer  that the output will be the
    // symbolic dimension value of dim1.

    // if `symbolic_shape_values` is not null, record list accesses
    // which resolve to symbolic dimension values with their concrete symbolic
    // shape value. Because symbolic dimensions are represented as negative numbers and
    // are not real values, inserting them as constants in the graph would invalidate
    // the graph for further use. Instead, we keep track of what their value would be
    // for extracting output shapes.
    // clang-format on

    std::unordered_map<int64_t, std::vector<Value*>> symbolic_shape_map;

    TORCH_INTERNAL_ASSERT(
        inputs_.size() >= shape_compute_graph_->inputs().size(),
        "Missing Arg for Shape Graph");
    for (const auto index :
         c10::irange(shape_compute_graph_->inputs().size())) {
      auto shape_arguments = std::get_if<ShapeArguments>(&inputs_[index]);
      if (!shape_arguments || !shape_arguments->has_dim()) {
        continue;
      }
      // Add support for testing symbolic shapes with dynamic dims

      for (const Use& use : shape_compute_graph_->inputs().at(index)->uses()) {
        // TODO: either decompose composite ops like slice or add handling here
        switch (use.user->kind()) {
          case aten::len: {
            size_t len = shape_arguments->len();
            replaceWithIValue(use.user->output(), static_cast<int64_t>(len));
          } break;
          case aten::__getitem__: {
            auto index = constant_as<int64_t>(use.user->inputs().at(1));
            if (!index) {
              continue;
            }
            auto norm_index = normIndex(*index, shape_arguments->len());
            if (!norm_index) {
              continue;
            }
            auto shape_arg = shape_arguments->at(*norm_index);
            if (auto const_int = shape_arg.asConstantInt()) {
              replaceWithIValue(use.user->output(), const_int);
              continue;
            }
            auto maybe_shape_symbol = shape_arg.asShapeSymbol();
            if (!maybe_shape_symbol) {
              continue;
            }
            auto shape_symbol = *maybe_shape_symbol;
            if (symbolic_shape_values) {
              symbolic_shape_values->emplace(
                  use.user->output(), shape_symbol.value());
            } else {
              int64_t symbolic_index = shape_symbol.value();
              symbolic_shape_map[symbolic_index].push_back(use.user->output());
            }
            for (const auto& sym_uses : use.user->output()->uses()) {
              auto k = sym_uses.user->kind();
              if (k != aten::ge && k != aten::le && k != aten::ne &&
                  k != aten::eq && k != aten::lt && k != aten::gt) {
                break;
              }
              auto other_index = 1 - sym_uses.offset;
              auto other_value =
                  constant_as<int64_t>(sym_uses.user->input(other_index));
              if (!other_value) {
                continue;
              }

              // check for dim >= 0, 0 <= dim
              // dim >= 0
              if (k == aten::ge && *other_value == 0 && other_index == 1) {
                replaceWithIValue(sym_uses.user->output(), true);
                continue;
              }
              // 0 <= dim
              if (k == aten::le && *other_value == 0 && other_index == 0) {
                replaceWithIValue(sym_uses.user->output(), true);
                continue;
              }

              // check for dim comparisons to negative number
              if (*other_value >= 0) {
                continue;
              }
              if (k == aten::eq || k == aten::ne) {
                // True if:
                // -2 != {Positive}
                replaceWithIValue(sym_uses.user->output(), k == aten::ne);
              } else {
                // True if:
                // -2 <= / < {Positive}
                // {Positive} >= / > {-2}
                bool true_val =
                    ((other_index == 0 && (k == aten::le || k == aten::lt)) ||
                     (other_index == 1 && (k == aten::ge || k == aten::gt)));
                replaceWithIValue(sym_uses.user->output(), true_val);
              }
            }
          }
        }
      }

      for (const auto& symbolic_set : symbolic_shape_map) {
        mergeSymbolicShapeSets(symbolic_set.second);
      }
    }
  }

  void mergeSymbolicShapeSets(const std::vector<Value*>& symbolic_set) {
    // `symbolic_set` represents a set of Value * which are all equal
    // to each other. Here, we optimize the graph by replacing values
    // in the set with other dominating values.
    // in the following example, where a, b and c are all in the same
    // symbolic set:
    // if cond:
    //    a = li[0]
    //    b = li[1]
    //    return [a, b]
    // else:
    //    c = li[0]
    //    return [c, c]
    // we can replace `b` with `a` because it is dominated by `a`,
    // but we cannot replace `c` with another dominating value

    // there are ways to compute this more efficiently but typically number of
    // Values for each symbolic set is low and this is cheap to run
    for (const auto i : c10::irange(symbolic_set.size())) {
      Value* v = symbolic_set[i];
      Value* dominating_value = v;
      for (const auto& sym_set : symbolic_set) {
        if (dominating_value->node()->isDominatedBy(sym_set->node())) {
          dominating_value = sym_set;
        }
      }
      if (dominating_value != v) {
        v->replaceAllUsesWith(dominating_value);
      }
    }
  }

  std::vector<c10::SymbolicShape> propagateShapesInGraph() {
    bool made_change = true;
    constexpr size_t MAX_ATTEMPTS = 8;
    for (unsigned attempt_num = 0; made_change && attempt_num < MAX_ATTEMPTS;
         attempt_num++) {
      // symbolic shape concrete values are only used in final shape extraction
      GRAPH_DUMP("Before substitution: ", shape_compute_graph_);
      substituteSymbolicProperties(/*symbolic_shape_values*/ nullptr);
      GRAPH_DUMP("Before Opt: ", shape_compute_graph_);
      made_change = shapeGraphCleanupPasses(shape_compute_graph_);
    }
    std::unordered_map<Value*, int64_t> symbolic_shape_values;
    substituteSymbolicProperties(&symbolic_shape_values);
    GRAPH_DUMP("Done with partial evaluation", shape_compute_graph_);

    return extractOutputShape(symbolic_shape_values);
  }

  std::vector<c10::SymbolicShape> extractOutputShape(
      std::unordered_map<Value*, int64_t>& symbolic_shape_values) {
    TORCH_INTERNAL_ASSERT(
        shape_compute_graph_->outputs().size() == schema_->returns().size());
    // TODO: would be nice if there were easy facility to look at uses and see
    // if they are all pure instead of instantiating db.
    auto res = std::vector<c10::SymbolicShape>();
    AliasDb db(shape_compute_graph_);
    for (size_t i = 0; i < shape_compute_graph_->outputs().size(); ++i) {
      auto output = shape_compute_graph_->outputs().at(i);
      auto type = output->type();
      TORCH_INTERNAL_ASSERT(isListOfInts(type));
      c10::SymbolicShape ss =
          extractListShape(output, symbolic_shape_values, db);
      GRAPH_DEBUG("Extracted Output: ", ss);
      res.push_back(ss);
    }
    return res;
  }

 public:
  SymbolicShapeOpAnalyzer(const FunctionSchema* schema) : schema_(schema) {
    shape_compute_graph_ = nullptr;
    if (!schema_) {
      return;
    }
    auto maybe_graph = shapeComputeGraphForSchema(*schema_);
    if (!maybe_graph) {
      return;
    }
    shape_compute_graph_ = (*maybe_graph)->copy();
  }

  SymbolicShapeOpAnalyzer(
      const FunctionSchema* schema,
      std::shared_ptr<Graph> graph)
      : schema_(schema) {
    shape_compute_graph_ = graph->copy();
  }

  std::optional<std::vector<c10::SymbolicShape>> run(
      std::vector<SSArgument>& inputs) {
    if (!shape_compute_graph_) {
      return std::nullopt;
    }
    inputs_ = inputs;
    substituteConstantInputs();
    GRAPH_DEBUG(inputs_)
    return propagateShapesInGraph();
  }

  std::shared_ptr<Graph> getShapeComputeGraph() {
    return shape_compute_graph_;
  }
};

SSArgument tensorShapeArg(Value* tensor_v) {
  auto tt = tensor_v->type()->expect<TensorType>();
  c10::SymbolicShape symbolic_shapes = tt->symbolic_sizes();

  // for testing, we don't insert complete tensor shapes and rely on our
  // partial evaluation pipeline to propagate information.
  // this is a good proxy for our ability to propagate non-complete shape
  // information.
  if (symbolic_shapes.isComplete() && !symbolic_shape_analysis_test_mode) {
    return IValue(tt->sizes().concrete_sizes());
  }
  if (toIValue(tensor_v)) {
    auto size = constant_as<at::Tensor>(tensor_v)->sizes();
    if (!symbolic_shape_analysis_test_mode) {
      return IValue(size);
    } else {
      return c10::SymbolicShape(size);
    }
  }
  return symbolic_shapes;
}

std::vector<SSArgument> getNodeInputShapes(Node* n, const AliasDb& db) {
  // TODO: fix the List of integers implementation, and
  // extract out the shape changes, otherwise this is complete
  // NB: shape compute graphs may have less inputs than their node
  // counterparts to allow e.g. sharing one single unary definition
  // so iterate on # of shape inputs
  // We make lists of Tensor inputs variadic, which results in
  // offset between a node index and its corresponding graph index
  std::vector<SSArgument> input_shapes = std::vector<SSArgument>();

  for (size_t node_index = 0; node_index < n->inputs().size(); ++node_index) {
    auto type = n->input(node_index)->type();

    if (type->castRaw<TensorType>()) {
      input_shapes.push_back(tensorShapeArg(n->input(node_index)));
      continue;
    }
    if (isListOfTensors(type)) {
      // waiting for more use cases to decide on best generalization
      if (n->input(node_index)->node()->kind() == prim::Constant) {
        auto ival = toIValue(n->input(node_index));
        for (const auto& ten : ival->toTensorVector()) {
          input_shapes.emplace_back(c10::List<int64_t>(ten.sizes()));
        }
      } else if (
          n->input(node_index)->node()->kind() == prim::ListConstruct &&
          !db.hasWriters(n->input(node_index))) {
        auto li_construct_node = n->input(node_index)->node();
        for (size_t j = 0; j < li_construct_node->inputs().size(); ++j) {
          input_shapes.push_back(tensorShapeArg(li_construct_node->input(j)));
        }
      } else {
        TORCH_INTERNAL_ASSERT(false, "Unhandled List, we shouldn't get here");
      }
      continue;
    }
    if (auto ival = toIValue(n->input(node_index))) {
      input_shapes.emplace_back(*ival);
      continue;
    }
    if (type->cast<ListType>() &&
        type->cast<ListType>()->getElementType()->cast<IntType>()) {
      auto input_src_node = n->input(node_index)->node();
      if (input_src_node->kind() == prim::ListConstruct &&
          !db.hasWriters(n->input(node_index))) {
        // it is a very common in graphs to see patterns like:
        // z = x.view(y.size())
        // or:
        // z = x.view(1, 10, y.size(0), y.size(1))
        // We want to propagate symbolic dimensions and concrete sizes
        // from y to z. To do this we try to associate symbolic dimensions
        // or concrete sizes with the integer list inputs that have a
        // constructor taken from constants or y.size() or y.size(0)
        auto list_construct = n->input(node_index)->node();
        std::vector<ShapeArg> shape;
        for (Value* v : list_construct->inputs()) {
          if (auto constant = constant_as<int64_t>(v)) {
            shape.emplace_back(*constant);
          } else if (v->node()->kind() == aten::size) {
            auto const_index = constant_as<int64_t>(v->node()->input(1));
            auto tt = v->node()->input(0)->type()->expect<TensorType>();
            auto ss = tt->symbolic_sizes();
            if (!ss.rank() || !const_index) {
              // if we are getting a size of a tensor, it is an unknown
              // symbolic dimension instead of an unknown integer (must be
              // >=0)
              shape.emplace_back(at::ShapeSymbol::newSymbol());
              continue;
            }
            auto norm_index = normIndex(*const_index, *ss.rank());
            if (!norm_index) {
              shape.emplace_back(at::ShapeSymbol::newSymbol());
              continue;
            }
            shape.emplace_back(ss[*norm_index]);
          } else {
            shape.emplace_back(ShapeArg::unknownInteger());
          }
        }
        input_shapes.emplace_back(ShapeArguments(shape));
        continue;
      }
      if (input_src_node->kind() == aten::size &&
          !db.hasWriters(n->input(node_index))) {
        auto ten_inp = input_src_node->input();
        auto ss = ten_inp->type()->expect<TensorType>()->symbolic_sizes();
        input_shapes.emplace_back(ss);
        continue;
      }
    }
    GRAPH_DEBUG(
        "Unhandled input: ",
        n->kind().toDisplayString(),
        " arg num: ",
        node_index);
    input_shapes.emplace_back(c10::SymbolicShape());
  }
  TORCH_INTERNAL_ASSERT(
      input_shapes.size() >= n->inputs().size(),
      "input_shapes size: ",
      input_shapes.size(),
      " n inputs size: ",
      n->inputs().size());
  return input_shapes;
}

void applyOutputShapeToGraph(
    Node* node,
    const std::vector<c10::SymbolicShape>& output_shapes) {
  TORCH_INTERNAL_ASSERT(
      node->outputs().size() == output_shapes.size(),
      "Output shape size mismatch");
  for (size_t i = 0; i < output_shapes.size(); ++i) {
    auto& ss = output_shapes.at(i);
    node->output(i)->setType(
        node->output(i)->type()->expect<TensorType>()->withSymbolicShapes(ss));
  }
}

std::shared_ptr<Graph> PropagateShapesWithShapeFunction(
    Node* n,
    const AliasDb& db) {
  const FunctionSchema* func_schema = n->maybeSchema();
  if (!func_schema) {
    return nullptr;
  }
  auto op_analyzer = SymbolicShapeOpAnalyzer(func_schema);
  if (!op_analyzer.getShapeComputeGraph()) {
    return nullptr;
  }
  auto input_shapes = getNodeInputShapes(n, db);
  op_analyzer.refineInputUnionTypes(n);

  if (auto output_shapes = op_analyzer.run(input_shapes)) {
    applyOutputShapeToGraph(n, *output_shapes);
  }

  return op_analyzer.getShapeComputeGraph();
}

c10::SymbolicShape combine_bounds(
    c10::SymbolicShape& lower_bound,
    c10::SymbolicShape& upper_bound) {
  // TODO: At some point we might want to add support for dynamic dims
  TORCH_INTERNAL_ASSERT(lower_bound.rank() == upper_bound.rank());
  if (lower_bound.rank() == std::nullopt) {
    return c10::SymbolicShape();
  }
  std::vector<c10::ShapeSymbol> merged_shapes;
  for (const auto i : c10::irange(*lower_bound.rank())) {
    // TODO: Merge equivalent expressions (not needed for current use case)
    if (lower_bound[i] == upper_bound[i]) {
      merged_shapes.push_back(lower_bound[i]);
    } else {
      merged_shapes.push_back(c10::ShapeSymbol::newSymbol());
    }
  }
  return c10::SymbolicShape(std::move(merged_shapes));
}

struct SymbolicShapeGraphAnalyzer {
  SymbolicShapeGraphAnalyzer(
      std::shared_ptr<Graph>& graph,
      Node* beg,
      Node* end)
      : graph_(graph), beg_(beg), end_(end) {
    TORCH_INTERNAL_ASSERT(
        beg_->owningBlock() == end_->owningBlock() && end_->isAfter(beg_));
  }

  std::optional<ShapeComputeGraphMapping> run() {
    AliasDb db(graph_);
    std::unordered_map<Node*, std::shared_ptr<Graph>> partial_evaluated_graphs =
        propagateShapesAndGatherPartialEvalShapeGraphs(db);

    auto stitched_shape_compute_graph = std::make_shared<Graph>();
    // We want to build up a computational graph which computes all shapes
    // we dont know statically - that is, all symbolic shapes within
    // the region [beg, end). it must be executable before beg.
    // TODO: dont require dimensions of tensors to be set AOT ?

    for (auto it = beg_->iterator(); it != end_->iterator(); it++) {
      auto curr = *it;
      if (curr->kind() == prim::Constant) {
        continue;
      }
      // TODO: generalize logic to for other tensor input ops when they are
      // added
      if (curr->kind() == prim::ListConstruct) {
        auto uses = curr->output()->uses();
        if (!std::all_of(uses.begin(), uses.end(), [](const Use& use) {
              return use.user->kind() == aten::cat;
            })) {
          GRAPH_DEBUG("Non cat list use ", getHeader(curr));
          return std::nullopt;
        }
        continue;
      }

      if (!partial_evaluated_graphs.count(curr)) {
        GRAPH_DEBUG("No graph ", getHeader(curr));
        return std::nullopt;
      }

      auto outputs = curr->outputs();
      for (Value* v : outputs) {
        auto tt = v->type()->cast<TensorType>();
        if (!tt) {
          GRAPH_DEBUG("Non tensor node", getHeader(curr));
          return std::nullopt;
        }
        auto symbolic_sizes = tt->symbolic_sizes();
        // TODO: dont require # of dimensions of tensors set ?
        if (!symbolic_sizes.rank()) {
          GRAPH_DEBUG("No rank on output ", getHeader(curr));
          return std::nullopt;
        }
      }
      auto partial_eval_graph = partial_evaluated_graphs[curr];
      joinPartialEvaluatedShapeGraphToLargeShapeGraph(
          curr, partial_eval_graph, stitched_shape_compute_graph);
    }

    size_t MAX_ITER = 8;
    bool made_change = true;
    size_t i = 0;
    while (i < MAX_ITER && made_change) {
      i++;
      made_change = shapeGraphCleanupPasses(stitched_shape_compute_graph);
    }

    // for any output that is duplicated, the symbolic shape must be equal
    // take the symbolic shape that is generated first and get equivalent ones
    std::unordered_map<int64_t, int64_t> discovered_sym_shape_equalities;
    std::unordered_map<Value*, int64_t> graph_output_to_symbolic_shape_dim;
    std::vector<size_t> erase_indices;

    for (size_t i = 0; i < stitched_shape_compute_graph->outputs().size();
         ++i) {
      Value* output = stitched_shape_compute_graph->outputs().at(i);
      // this Value is already contained, so the symbolic shape for i must be
      // equal to the symbolic shape at the existing index
      if (graph_output_to_symbolic_shape_dim.count(output)) {
        auto curr_sym_shape = output_index_to_symbolic_shape_[i];
        auto existing_sym_shape = graph_output_to_symbolic_shape_dim[output];
        discovered_sym_shape_equalities[curr_sym_shape] = existing_sym_shape;
        erase_indices.push_back(i);
      } else {
        graph_output_to_symbolic_shape_dim[output] =
            output_index_to_symbolic_shape_[i];
      }
    }
    for (int64_t i = erase_indices.size() - 1; i >= 0; i--) {
      stitched_shape_compute_graph->eraseOutput(erase_indices[i]);
    }
    for (size_t i = 0; i < stitched_shape_compute_graph->inputs().size();) {
      if (!stitched_shape_compute_graph->inputs().at(i)->hasUses()) {
        enclosing_graph_value_to_shape_graph_input_.erase(
            stitched_shape_compute_graph->inputs().at(i));
        stitched_shape_compute_graph->eraseInput(i);
      } else {
        ++i;
      }
    }

    updateGraphWithSymbolicShapeEqualities(discovered_sym_shape_equalities);
    return ShapeComputeGraphMapping(
        std::move(stitched_shape_compute_graph),
        enclosing_graph_value_to_shape_graph_input_,
        std::move(graph_output_to_symbolic_shape_dim));
  }

  void updateGraphWithSymbolicShapeEqualities(
      std::unordered_map<int64_t, int64_t>& sym_shape_equalities) {
    for (auto it = beg_->iterator(); it != end_->iterator(); it++) {
      auto curr = *it;
      for (size_t i = 0; i < curr->outputs().size(); ++i) {
        auto output = curr->output(i);
        auto tt = output->type()->cast<TensorType>();
        if (!tt || !tt->symbolic_sizes().rank()) {
          continue;
        }
        bool changed = false;
        std::vector<at::ShapeSymbol> shape_vec = *tt->symbolic_sizes().sizes();
        auto new_sizes =
            c10::fmap(shape_vec, [&](const at::ShapeSymbol& shape) {
              auto value = shape.value();
              if (sym_shape_equalities.count(value)) {
                changed = true;
                return sym_shape_equalities[value];
              }
              return value;
            });
        if (changed) {
          output->setType(
              tt->withSymbolicShapes(c10::SymbolicShape(new_sizes)));
        }
      }
    }
  }

  void registerStitchedComputeOutput(
      std::shared_ptr<Graph> stitched_shape_compute_graph,
      Value* output,
      int64_t symbolic_shape) {
    stitched_shape_compute_graph->registerOutput(output);
    output_index_to_symbolic_shape_
        [stitched_shape_compute_graph->outputs().size() - 1] = symbolic_shape;
    symbolic_shape_value_to_graph_output_[symbolic_shape] =
        stitched_shape_compute_graph->outputs().at(
            stitched_shape_compute_graph->outputs().size() - 1);
  }

  void joinPartialEvaluatedShapeGraphToLargeShapeGraph(
      Node* curr,
      std::shared_ptr<Graph> partial_eval_graph,
      std::shared_ptr<Graph> stitched_shape_compute_graph) {
    // we are building up the large shape compute graph by iteratively
    // combining partially evaluated individual node shape graphs.

    // We need to maintain two mappings, one from non-Tensor inputs in the
    // enclosing graph to their equivalent mappings within the large shape
    // compute graph, and one from symbolic shape dimension to new node output

    // When we add a new tensor node, we do two things:
    // 1: record a mapping from the tensor node output to its shape in the
    // partial eval graph 2: add each symbolic shape dimension that we have
    // not already added as a output to the large shape compute graph

    // Once we are done stitching together all partial eval'd graphs, we can
    // cleanup the graph and remove the unneeded complete shapes as outputs,
    // leaving us only compute for calculating the runtime value of symbolic
    // dimensions
    // leaving us only compute for calculating the runtime value of symbolic
    // dimensions

    std::vector<Value*> node_inputs;
    // TODO: generalize logic
    if (curr->kind() == aten::cat) {
      TORCH_INTERNAL_ASSERT(
          curr->input(0)->node()->kind() == prim::ListConstruct);
      for (Value* v : curr->input(0)->node()->inputs()) {
        node_inputs.push_back(v);
      }
      node_inputs.push_back(curr->namedInput("dim"));
    } else {
      for (size_t i = 0; i < partial_eval_graph->inputs().size(); ++i) {
        node_inputs.push_back(curr->input(i));
      }
    }

    std::vector<Value*> partial_eval_inputs;
    for (size_t i = 0; i < node_inputs.size(); ++i) {
      auto node_input = node_inputs[i];
      auto existing_graph_mapping =
          enclosing_graph_value_to_shape_graph_input_.find(node_input);
      if (existing_graph_mapping !=
          enclosing_graph_value_to_shape_graph_input_.end()) {
        partial_eval_inputs.push_back(existing_graph_mapping->second);
      } else {
        Value* shape_graph_input =
            stitched_shape_compute_graph->addInput()->copyMetadata(
                partial_eval_graph->inputs().at(i));
        enclosing_graph_value_to_shape_graph_input_[node_input] =
            shape_graph_input;
        partial_eval_inputs.push_back(shape_graph_input);
      }
      // make sure all symbolic dimensions in the graph we are creating are
      // computed in the partial eval graph
      if (auto tt = node_input->type()->cast<TensorType>()) {
        if (!tt->symbolic_sizes().rank()) {
          continue;
        }
        auto rank = *tt->symbolic_sizes().rank();
        for (size_t j = 0; j < rank; ++j) {
          auto shape = tt->symbolic_sizes()[j];
          if (shape.is_static() ||
              symbolic_shape_value_to_graph_output_.count(shape.value())) {
            continue;
          }
          auto input = enclosing_graph_value_to_shape_graph_input_[node_input];
          WithInsertPoint guard(stitched_shape_compute_graph->block());
          auto index = stitched_shape_compute_graph->insertConstant(
              static_cast<int64_t>(j));
          auto li_index = stitched_shape_compute_graph->insert(
              aten::__getitem__, {input, index});
          registerStitchedComputeOutput(
              stitched_shape_compute_graph, li_index, shape.value());
        }
      }
    }

    WithInsertPoint guard(stitched_shape_compute_graph->block());
    std::unordered_map<Value*, Value*> value_map;
    insertGraph(
        *stitched_shape_compute_graph,
        *partial_eval_graph,
        partial_eval_inputs,
        value_map);

    for (size_t i = 0; i < curr->outputs().size(); ++i) {
      Value* new_list_output = value_map[partial_eval_graph->outputs().at(i)];
      enclosing_graph_value_to_shape_graph_input_[curr->output(i)] =
          new_list_output;

      TORCH_INTERNAL_ASSERT(
          new_list_output->node()->kind() == prim::ListConstruct ||
          new_list_output->node()->kind() == prim::Constant);
      TORCH_INTERNAL_ASSERT(!new_list_output->node()->hasUses());

      auto symbolic_sizes =
          curr->output(i)->type()->expect<TensorType>()->symbolic_sizes();
      TORCH_INTERNAL_ASSERT(symbolic_sizes.rank());

      for (size_t i = 0; i < *symbolic_sizes.rank(); i++) {
        if (symbolic_sizes[i].is_static()) {
          continue;
        }
        int64_t symbolic_shape = symbolic_sizes[i].value();
        if (symbolic_shape_value_to_graph_output_.count(symbolic_shape)) {
          continue;
        }
        registerStitchedComputeOutput(
            stitched_shape_compute_graph,
            new_list_output->node()->input(i),
            symbolic_shape);
      }
    }
  }

  std::unordered_map<Node*, std::shared_ptr<Graph>>
  propagateShapesAndGatherPartialEvalShapeGraphs(AliasDb& db) {
    std::unordered_map<Node*, std::shared_ptr<Graph>> partial_evaluated_graphs;
    for (auto it = beg_->iterator(); it != end_->iterator(); it++) {
      auto curr = *it;
      if (auto maybe_graph = PropagateShapesWithShapeFunction(curr, db)) {
        partial_evaluated_graphs[curr] = maybe_graph;
      }
    }
    return partial_evaluated_graphs;
  }

  std::unordered_map<Value*, Value*>
      enclosing_graph_value_to_shape_graph_input_;
  std::unordered_map<int64_t, Value*> symbolic_shape_value_to_graph_output_;
  std::unordered_map<size_t, int64_t> output_index_to_symbolic_shape_;

  std::shared_ptr<Graph>& graph_;
  Node* beg_;
  Node* end_;
};

void PropagateShapesOnBlock(Block* b, const AliasDb& db) {
  for (Node* n : b->nodes()) {
    // TODO: handle loop
    if (n->kind() == prim::If) {
      IfView if_v(n);
      PropagateShapesOnBlock(if_v.thenBlock(), db);
      PropagateShapesOnBlock(if_v.elseBlock(), db);
      mergeTypes(if_v.thenOutputs(), if_v.elseOutputs(), if_v.outputs());
    } else if (n->maybeSchema()) {
      PropagateShapesWithShapeFunction(n, db);
    } else if (n->kind() == prim::TupleConstruct) {
      auto orig_type = n->output()->type()->expect<TupleType>();
      auto new_types = fmap(n->inputs(), [](Value* v) { return v->type(); });
      n->output()->setType(
          orig_type->createWithContained(std::move(new_types)));
    }
  }
}
} // namespace

void PropagateShapesOnGraph(std::shared_ptr<Graph>& graph) {
  AliasDb db(graph);
  PropagateShapesOnBlock(graph->block(), db);
}

std::optional<ShapeComputeGraphMapping>
PropagateShapesAndBuildLargeShapeComputeGraph(
    std::shared_ptr<Graph>& graph,
    Node* beg,
    Node* end) {
  return SymbolicShapeGraphAnalyzer(graph, beg, end).run();
}

TORCH_API std::optional<std::vector<c10::SymbolicShape>>
calculateSymbolicShapesOnOp(
    const FunctionSchema* schema,
    const std::vector<SSAInput>& inputs) {
  auto bounded_graphs = boundedGraphsForSchema(*schema);
  auto has_shape_compute = shapeComputeGraphForSchema(*schema) != std::nullopt;
  if (!has_shape_compute && bounded_graphs == std::nullopt) {
    // Avoid doing all this work for functions that don't have a
    // supported schema
    return std::nullopt;
  }

  if (auto cached_ret_vec = get_cached_shape_function(schema, inputs)) {
    return cached_ret_vec;
  }

  std::vector<SSArgument> ssa_args;
  for (auto& arg : inputs) {
    if (const IValue* ival = std::get_if<IValue>(&arg)) {
      ssa_args.emplace_back(*ival);
    } else {
      const c10::SymbolicShape* ss = std::get_if<c10::SymbolicShape>(&arg);
      ssa_args.emplace_back(ShapeArguments(*ss));
    }
  }
  // Handle bounded shape option
  if (bounded_graphs) {
    auto lower_bound =
        SymbolicShapeOpAnalyzer(schema, bounded_graphs->lower_bound);
    auto lower_bound_res = lower_bound.run(ssa_args);
    auto upper_bound =
        SymbolicShapeOpAnalyzer(schema, bounded_graphs->upper_bound);
    auto upper_bound_res = upper_bound.run(ssa_args);
    // Stitch together the values
    if (lower_bound_res.has_value() && upper_bound_res.has_value()) {
      TORCH_INTERNAL_ASSERT(lower_bound_res->size() == upper_bound_res->size());
      auto merged_res = std::vector<c10::SymbolicShape>();
      for (size_t i = 0; i < lower_bound_res->size(); i++) {
        merged_res.push_back(
            combine_bounds(lower_bound_res->at(i), upper_bound_res->at(i)));
      }
      cache_shape_function(schema, inputs, merged_res);
      return merged_res;
    }
    return std::nullopt;
  }

  auto op_analyzer = SymbolicShapeOpAnalyzer(schema);
  auto res = op_analyzer.run(ssa_args);
  if (res.has_value()) {
    cache_shape_function(schema, inputs, res.value());
  }
  return res;
}

} // namespace jit
} // namespace torch