torch/csrc/jit/passes/quantization.cpp

#include <torch/csrc/jit/passes/quantization.h>
#include <torch/csrc/jit/passes/subgraph_rewrite.h>
#include <torch/csrc/jit/passes/constant_propagation.h>

#include <torch/csrc/jit/ir.h>
#include <torch/csrc/jit/irparser.h>
#include <torch/csrc/jit/jit_log.h>
#include <torch/csrc/jit/node_hashing.h>
#include <torch/csrc/jit/operator.h>
#include <torch/csrc/jit/subgraph_matcher.h>

#include <stack>

namespace torch {
namespace jit {
namespace {

void findValuesInPattern(
    Graph& graph,
    const std::string& pattern,
    std::unordered_set<Value*>& values_to_skip) {
  Graph pattern_graph;
  std::unordered_map<std::string, Value*> vmap;
  script::parseIR(pattern, &pattern_graph, vmap);

  auto matches = findPatternMatches(pattern_graph, graph);
  for (const auto& match : matches) {
    auto output_value = vmap.at("output");
    TORCH_INTERNAL_ASSERT(
        match.values_map.find(output_value) != match.values_map.end(),
        "Didn't find Value output in match result.");
    values_to_skip.emplace(match.values_map.at(output_value));
  }
}

void addIntermediateValuesToSkipObserver(
    const script::Module& module,
    const std::string& method_name,
    std::unordered_set<Value*>& values_to_skip) {
  script::Method method = module.get_method(method_name);
  auto graph = method.graph();

  // Note that the name of the value we want to skip inserting observer for
  // is hard coded as "output"
  std::string conv_functional_relu = R"(
graph(%self, %input, %inplace):
    %relu = prim::Constant[name="relu"]()
    %conv = match::module[name="Conv2d"](%self)
    %output = prim::CallMethod[name="forward"](%conv, %input)
    %r = prim::CallFunction(%relu, %output, %inplace)
    return (%r))";
  std::string conv_relu_module = R"(
graph(%self, %input):
    %conv = match::module[name="Conv2d"](%self)
    %output = prim::CallMethod[name="forward"](%conv, %input)
    %relu = match::module[name="ReLU"](%self)
    %r = prim::CallMethod[name="forward"](%relu, %output)
    return (%r))";
  std::vector<std::string> patterns = {conv_functional_relu, conv_relu_module};

  for (const auto& pattern : patterns) {
    findValuesInPattern(*graph, pattern, values_to_skip);
  }
}

static bool outputsNeedToBeObserved(Node* n) {
  return n->kind() != prim::Constant;
}

Node* traverseToQuantNode(Node* dq) {
  TORCH_INTERNAL_ASSERT(dq != nullptr);
  TORCH_INTERNAL_ASSERT(dq->inputs().size() != 0);
  Node* intrepr = dq->inputs()[0]->node();
  TORCH_INTERNAL_ASSERT(intrepr != nullptr);
  TORCH_INTERNAL_ASSERT(intrepr->inputs().size() != 0);
  return intrepr->inputs()[0]->node();
}

Value* insertScalarType(Node* ins_node, at::ScalarType t) {
  TORCH_INTERNAL_ASSERT(t != at::ScalarType::Undefined);
  WithInsertPoint ins(ins_node);
  // ScalarType inserted before ins_node node which is
  // beginning of the quant-dequant pattern
  Value* scalartype_v =
      ins_node->owningGraph()->insertConstant(IValue(static_cast<int>(t)));
  return scalartype_v;
}

// Create Quant Node
Node* createQuantNode(Value* v, Graph* g) {
  Node* quant = g->create(at::Symbol::fromQualString("aten::quantize_linear"));
  TORCH_INTERNAL_ASSERT(quant != nullptr, "Failed to create quant node");
  quant->output()->setDebugName(v->debugName() + ".quant");
  return quant;
}

// Create Dequant node
Node* createDeQuantNode(Value* v, Graph* g) {
  Node* dequant =
      g->create(at::Symbol::fromQualString("aten::_dequantize_linear"));
  TORCH_INTERNAL_ASSERT(dequant != nullptr, "Failed to create dequant node");
  dequant->output()->setDebugName(v->debugName() + ".dequant");
  return dequant;
}

// Create IntTensor Node
Node* createIntReprNode(Value* v, Graph* g) {
  Node* intrepr = g->create(at::Symbol::fromQualString("aten::int_repr"));
  TORCH_INTERNAL_ASSERT(intrepr != nullptr, "Failed to create inttensor node");
  intrepr->output()->setDebugName(v->debugName() + ".intrepr");
  return intrepr;
}

// Clone observer module and add it to the original module,
// and insert a call to observer forward function
Node* insertObserver(
    Value* v,
    Graph* g,
    script::Module& module,
    const QConfig& qconfig) {
  script::Module observer_module;
  if (v->node()->kind() == prim::GetAttr &&
      v->node()->s(attr::name) == "weight") {
    observer_module = std::get<1>(qconfig);
  } else {
    observer_module = std::get<0>(qconfig);
  }
  std::string observer_name = "observer_for_" + v->debugName();
  // Temporary workaround to skip inserting duplicate modules,
  // full support will come in next PR
  for (script::Slot s: module.get_module_slots()) {
    if (s.name() == observer_name) {
      return nullptr;
    }
  }
  script::Module observer = observer_module.clone();
  module.register_module(observer_name, observer);
  // Get handle of observer module
  Node* observer_instance = g->create(c10::prim::GetAttr);
  // self.observer_for_v
  observer_instance->addInput(g->inputs()[0]);
  observer_instance->s_(c10::attr::name, observer_name);
  observer_instance->output()->setDebugName(observer_name);
  observer_instance->output()->setType(observer.type());
  observer_instance->insertAfter(v->node());

  // Create forward method call
  Node* call = g->create(c10::prim::CallMethod);
  TORCH_INTERNAL_ASSERT(call != nullptr, "Failed to create forward call node");
  call->s_(c10::attr::name, "forward");
  call->addInput(observer_instance->output());
  call->addInput(v);
  call->output()->setType(v->type());
  call->output()->setDebugName(v->debugName() + ".observed");
  call->insertAfter(observer_instance);
  return call;
}

void fillQConfigMap(
    const script::Module& module,
    const QConfigDict& qconfig_dict,
    ModuleQConfigMap& map,
    const std::string& key = "",
    const c10::optional<QConfig>& parent_qconfig = c10::nullopt) {
  c10::optional<QConfig> qconfig;
  if (qconfig_dict.find(key) != qconfig_dict.end()) {
    qconfig = qconfig_dict.at(key);
  } else {
    qconfig = parent_qconfig;
  }
  map[module.module_object()] = qconfig;

  for (script::Slot s : module.get_module_slots()) {
    std::string child_key;
    if (key == "") {
      child_key = s.name();
    } else {
      child_key = key + "." + s.name();
    }
    fillQConfigMap(s.to_module(), qconfig_dict, map, child_key, qconfig);
  }
}

void InsertObserversImpl(
    script::Module& module,
    const std::string& method_name,
    const ModuleQConfigMap& module_qconfig_map,
    std::unordered_set<Value*>& values_to_skip) {
  if (module_qconfig_map.count(module.module_object()) == 0) {
    // the module is added by us, e.g.: observer module
    return;
  }

  script::Method method = module.get_method(method_name);
  auto graph = method.graph();
  ConstantPropagation(graph);
  addIntermediateValuesToSkipObserver(module, method_name, values_to_skip);
  // For storing all values that need to be instrumented with an observer call.
  std::vector<Value*> values_to_observe;

  // For traversing all blocks in the graph including subblocks.
  std::stack<Block*> blocks_to_visit;

  // Mark observer nodes for inputs so we dont add observers
  // for observers while traversing graph
  std::unordered_set<Node*> observer_for_input;

  // Add observer for external input nodes excluding parameters
  // These are treated as activation as they vary across batches
  // and need to be observed.

  // prim::Param nodes do not belong to the graph. Hence the Insert
  // point is the beginning of graph node. This also safe guards against
  // observing a potentially mutated value due to some in-place operation
  Value* self = graph->inputs()[0];
  for (size_t idx = 1; idx < method.num_inputs(); ++idx) {
    auto& v = graph->inputs()[idx];
    if (v->type()->isSubtypeOf(TensorType::get()) &&
        values_to_skip.count(v) == 0) {
      auto qconfig = module_qconfig_map.at(module.module_object());
      if (qconfig) {
        auto observer_node =
            insertObserver(v, v->owningGraph(), module, qconfig.value());
        if (observer_node) {
          observer_for_input.emplace(observer_node);
        }
      }
    }
  }

  blocks_to_visit.push(graph->block());
  while (!blocks_to_visit.empty()) {
    Block* b = blocks_to_visit.top();
    blocks_to_visit.pop();
    for (Node* n : b->nodes()) {
      // Skip nodes that we don't need to observe, e.g. 'prim::Constant' or
      // observer nodes
      if (!outputsNeedToBeObserved(n) || observer_for_input.count(n) != 0) {
        continue;
      }

      // Record all outputs in the values_to_observe - we'll later add observers
      // for all values from it.
      for (Value* v : n->outputs()) {
        if (values_to_skip.count(v) == 0) {
          values_to_observe.push_back(v);
        }
        if (v->node()->kind() == prim::CallMethod) {
          // If we find a call to a method of a child module,
          // we'll recursively insert observers for the forward function to
          // the child module.
          auto module_instance = v->node()->inputs()[0];
          auto module_method_name = v->node()->s(attr::name);
          if (module_instance->node()->kind() == prim::GetAttr) {
            auto child_module_name = module_instance->node()->s(attr::name);
            auto child_module = module.find_module(child_module_name);
            TORCH_INTERNAL_ASSERT(
                child_module,
                "Child module " + child_module_name + " does not exist");
            // Recursively insert observer for the forward function of child module
            InsertObserversImpl(child_module.value(), module_method_name, module_qconfig_map, values_to_skip);
          } else {
            TORCH_INTERNAL_ASSERT(
                module_instance == graph->inputs()[0],
                "We only support call method either on %self"
                "or child instance in insert_observers_pass right now");
            InsertObserversImpl(module, module_method_name, module_qconfig_map, values_to_skip);
          }
        }
      }

      for (Block* subblock : n->blocks()) {
        blocks_to_visit.push(subblock);
      }
    }
  }

  // Actually add observer nodes.
  for (Value* v : values_to_observe) {
    if (!v->type()->isSubtypeOf(TensorType::get())) {
      continue;
    }
    // Skip inserting observer for bias
    if (v->node()->kind() == prim::GetAttr &&
        v->node()->s(c10::attr::name) == "bias") {
      continue;
    }
    auto qconfig = module_qconfig_map.at(module.module_object());
    // Skip inserting observer if no qconfig is specified
    if (qconfig) {
      insertObserver(v, v->owningGraph(), module, qconfig.value());
    }
  }
}

Node* insertQuantDeQuantCall(
    Value* v,
    const IValue& qparams,
    at::ScalarType t,
    bool insert_after = true) {
  Graph* g = v->node()->owningGraph();
  Node* quant = createQuantNode(v, g);
  Node* intrepr = createIntReprNode(v, g);
  Node* dequant = createDeQuantNode(v, g);
  Node* insert_point = insert_after ? v->node() : *g->nodes().begin();
  WithCurrentScope scope_guard(
      *insert_point->owningGraph(), insert_point->scope());
  WithInsertPoint ins(insert_point);

  // Add quant-intrepr-dequant nodes and replace for all uses of Value
  // Create qparam constant nodes
  TORCH_INTERNAL_ASSERT(qparams.isTuple(), "qparams must be tuple");
  auto tp = qparams.toTuple();
  IValue scale = tp->elements()[0].toTensor().item().toFloat();
  IValue zero_point = tp->elements()[1].toTensor().item().toInt();
  Value* scale_val = g->insertConstant(scale);
  Value* zero_point_val = g->insertConstant(zero_point);

  // Insert quant/int_repr/dequant nodes
  if (insert_after) {
    quant->insertAfter(insert_point);
  } else {
    quant->insertBefore(insert_point);
  }

  intrepr->insertAfter(quant);
  dequant->insertAfter(intrepr);

  // Attach inputs to quantization pattern nodes
  quant->addInput(v);
  intrepr->addInput(quant->output());
  dequant->addInput(intrepr->output());

  quant->addInput(scale_val);
  quant->addInput(zero_point_val);
  dequant->addInput(scale_val);
  dequant->addInput(zero_point_val);

  Value* scalar_type_val = insertScalarType(quant, t);
  TORCH_INTERNAL_ASSERT(scalar_type_val != nullptr);
  quant->addInput(scalar_type_val);
  dequant->addInput(scalar_type_val);
  return dequant;
}

// find the observer for Value `v` and return the name of the observer
c10::optional<std::string> findObserverName(Value* v) {
  for (const Use& u : v->uses()) {
    // Note that here we just check for the name of observer, but the ideally
    // we should be comparing the type of observer, this is a temporary
    // work around until data only clone of module.clone is supported.
    if (u.user->kind() == prim::CallMethod &&
        u.user->s(attr::name) == "forward") {
      auto module_instance = u.user->inputs().at(0);
      if (module_instance->node()->kind() == prim::GetAttr &&
          module_instance->node()->s(attr::name).find("observer_for_") !=
              std::string::npos) {
        return module_instance->node()->s(attr::name);
      }
    }
  }
  return c10::nullopt;
}

class QuantizeHelper {
 public:
  QuantizeHelper(const script::Module& m) : module_(m) {}
  IValue getQParams(Value* v);
  c10::optional<script::Module> findChildModuleToQuantize(Value* child_instance);
  void quantizeBias(Value* v);
  void quantizeTensor(Value* v, bool insert_after = true);
  void removeObserver(Value* v, const std::string& observer_name);
  void destroyNodes() {
    // Destroy observer forward calls
    for (auto& n : nodes_to_destroy_) {
      n->destroy();
    }
  }

 private:
  const script::Module& module_;
  std::vector<std::string> observer_modules_to_remove_;
  std::vector<Node*> nodes_to_destroy_;
};

void QuantizeHelper::removeObserver(
    Value* v,
    const std::string& observer_name) {
  // remove observer_module
  observer_modules_to_remove_.push_back(observer_name);
  // remove observer forward call
  for (const Use& u : v->uses()) {
    Node* user = u.user;
    if (user->kind() == prim::CallMethod && user->s(attr::name) == "forward" &&
        user->inputs()[0]->node()->kind() == prim::GetAttr &&
        user->inputs()[0]->node()->s(attr::name) == observer_name) {
      // Observer forward call node
      nodes_to_destroy_.push_back(user);
      // GetAttr node for observer module
      nodes_to_destroy_.push_back(user->inputs()[0]->node());
    }
  }
}

IValue QuantizeHelper::getQParams(Value* v) {
  TORCH_INTERNAL_ASSERT(v->type()->isSubtypeOf(TensorType::get()));
  auto observer_name = findObserverName(v);
  TORCH_INTERNAL_ASSERT(
      observer_name,
      "getQParams expects the corresponding observer for ",
      v->debugName(),
      " exists.");
  auto observer_module = module_.find_module(observer_name.value());
  TORCH_INTERNAL_ASSERT(
      observer_module,
      "getQParams expects the corresponding observer for ",
      v->debugName(),
      " exists.");
  auto calculate_qparams =
      observer_module.value().get_method("calculate_qparams");
  IValue qparams = calculate_qparams(std::vector<IValue>());
  return qparams;
}

double getScale(const IValue& qparam) {
  return qparam.toTuple()->elements()[0].toTensor().item().toDouble();
}

void QuantizeHelper::quantizeBias(Value* v) {
  // Traverse to the place where this is used
  std::vector<Symbol> ops_with_bias = {Symbol::aten("conv2d"),
                                       Symbol::aten("_convolution")};
  for (const auto& use : v->uses()) {
    if (std::find(
            ops_with_bias.begin(), ops_with_bias.end(), use.user->kind()) !=
        ops_with_bias.end()) {
      // Make sure there is no observer module for bias
      auto observer_name = findObserverName(v);
      TORCH_INTERNAL_ASSERT(!observer_name, "bias should not be observed!");
      Value* activation = use.user->inputs()[0];
      Value* weight = use.user->inputs()[1];
      // Get qparam from activation
      IValue act_qparam = getQParams(activation);
      // Get qparam from weight
      IValue weight_qparam = getQParams(weight);
      IValue bias_scale = at::scalar_tensor(
          c10::Scalar(getScale(act_qparam) * getScale(weight_qparam)),
          at::kDouble);
      IValue bias_qparam = c10::ivalue::Tuple::create(
          std::vector<IValue>({bias_scale, at::scalar_tensor(c10::Scalar(0))}),
          act_qparam.toTuple()->type);
      Node* dequant = insertQuantDeQuantCall(v, bias_qparam, at::kQInt32);
      v->replaceAllUsesWith(dequant->output());
      Node* q = traverseToQuantNode(dequant);
      TORCH_INTERNAL_ASSERT(q != nullptr);
      q->replaceInputWith(dequant->output(), v);
    }
  }
}

void QuantizeHelper::quantizeTensor(Value* v, bool insert_after) {
  auto observer_name = findObserverName(v);
  if (!observer_name) {
    return;
  }
  IValue qparams = getQParams(v);
  removeObserver(v, observer_name.value());
  Node* dequant;
  if (v->node()->kind() == prim::GetAttr &&
      v->node()->s(c10::attr::name) == "weight") {
    dequant = insertQuantDeQuantCall(v, qparams, at::kQInt8);
  } else {
    dequant = insertQuantDeQuantCall(v, qparams, at::kQUInt8, insert_after);
  }
  v->replaceAllUsesWith(dequant->output());
  Node* q = traverseToQuantNode(dequant);
  TORCH_INTERNAL_ASSERT(q);
  q->replaceInputWith(dequant->output(), v);
}

c10::optional<script::Module> QuantizeHelper::findChildModuleToQuantize(
    Value* child_instance) {
  TORCH_INTERNAL_ASSERT(
      child_instance->node()->kind() == prim::GetAttr,
      "Child instance should come from GetAttr.");
  auto child_module_name = child_instance->node()->s(attr::name);
  if (child_module_name.find("observer_for_") == std::string::npos) {
    auto child_module = module_.find_module(child_module_name);
    TORCH_INTERNAL_ASSERT(
        child_module,
        "InsertQuantDeQuant - Child module " + child_module_name +
        " does not exist");
    return child_module;
  }
  return c10::nullopt;
}

void InsertQuantDeQuantImpl(
    script::Module& module,
    const std::string& method_name) {
  script::Method method = module.get_method(method_name);
  auto graph = method.graph();

  // prim::Param nodes do not belong to the graph. Hence the Insert
  // point is the beginning of graph node. This also safe guards against
  // observing a potentially mutated value due to some in-place operation
  std::vector<Value*> input_values;
  for (size_t idx = 1; idx < method.num_inputs(); ++idx) {
    auto& v = graph->inputs()[idx];
    if (v->type()->isSubtypeOf(TensorType::get())) {
      input_values.push_back(v);
    }
  }

  QuantizeHelper qh(module);
  std::stack<Block*> blocks_to_visit;
  blocks_to_visit.push(graph->block());
  while (!blocks_to_visit.empty()) {
    Block* b = blocks_to_visit.top();
    blocks_to_visit.pop();
    for (auto it = b->nodes().begin(), end = b->nodes().end(); it != end;) {
      Node* n = *it++;
      for (Value* v : n->outputs()) {
        if (!v->type()->isSubtypeOf(TensorType::get())) {
          continue;
        }
        if (v->node()->kind() == prim::CallMethod) {
          auto module_instance = v->node()->inputs()[0];
          auto module_method_name = v->node()->s(attr::name);
          c10::optional<script::Module> m;
          // calling method on self
          if (module_instance == graph->inputs()[0]) {
            m = module;
          } else {
            m = qh.findChildModuleToQuantize(module_instance);
          }
          if (m) {
            InsertQuantDeQuantImpl(m.value(), module_method_name);
          }
        }
        if (v->node()->kind() == prim::GetAttr &&
            v->node()->s(c10::attr::name) == "bias") {
          qh.quantizeBias(v);
        } else {
          qh.quantizeTensor(v);
        }
      }

      for (Block* subblock : n->blocks()) {
        blocks_to_visit.push(subblock);
      }
    }
  }

  for (Value* v : input_values) {
    qh.quantizeTensor(v, false);
  }

  qh.destroyNodes();
}

} // namespace

TORCH_API void InsertObservers(
    script::Module& module,
    const std::string& method_name,
    const QConfigDict& qconfig_dict) {
  ModuleQConfigMap module_qconfig_map;
  fillQConfigMap(module,
                 qconfig_dict,
                 module_qconfig_map);
  std::unordered_set<Value*> values_to_skip;
  InsertObserversImpl(module,
                      method_name,
                      module_qconfig_map,
                      values_to_skip);
}

script::Module InsertQuantDeQuant(
    script::Module& input_module,
    const std::string& method_name) {
  script::Module module = input_module.clone();
  InsertQuantDeQuantImpl(module, method_name);

  // NOTE: Remove observer module does not work right now, we'll return
  // the module with observer modules as a temporary workaround
  // TODO: remove observer modules after we have a remove_module API
  return module;
}

// PyBind APIs
void PropagateQuantInfo(std::shared_ptr<Graph>& graph) {
  throw std::runtime_error("Pass not implemented yet!");
}

void QuantLinting(std::shared_ptr<Graph>& graph) {
  throw std::runtime_error("Pass not implemented yet!");
}

void FoldQuantNodesIntoInputsOutputs(std::shared_ptr<Graph>& graph) {
  throw std::runtime_error("Pass not implemented yet!");
}

void QuantFusion(std::shared_ptr<Graph>& graph) {
  std::string pattern = R"(
graph(%a_quant, %w_quant, %b_quant, %a_scale, %a_zero_point, %a_dtype, %w_scale, %w_zero_point, %w_dtype, %b_scale, %b_zero_point, %b_dtype, %r_scale, %r_zero_point, %r_dtype, %c, %d, %e, %f):
        %a_intrepr = aten::int_repr(%a_quant)
        %a_dequant = aten::_dequantize_linear(%a_intrepr, %a_scale, %a_zero_point, %a_dtype)
        %w_intrepr = aten::int_repr(%w_quant)
        %w_dequant = aten::_dequantize_linear(%w_intrepr, %w_scale, %w_zero_point, %w_dtype)
        %b_intrepr = aten::int_repr(%b_quant)
        %b_dequant = aten::_dequantize_linear(%b_intrepr, %b_scale, %b_zero_point, %b_dtype)
        %r = aten::conv2d(%a_dequant, %w_dequant, %b_dequant, %c, %d, %e, %f)
        %r_quant = aten::quantize_linear(%r, %r_scale, %r_zero_point, %r_dtype)
        return (%r_quant))";

  std::string replacement = R"(
graph(%a_quant, %w_quant, %b_quant, %a_scale, %a_zero_point, %a_dtype, %w_scale, %w_zero_point, %w_dtype, %b_scale, %b_zero_point, %b_dtype, %r_scale, %r_zero_point, %r_dtype, %stride, %padding, %dilation, %groups):
        %0 : int = prim::Constant[value=0]()
        %1 : int = prim::Constant[value=1]()
        %2 : int = prim::Constant[value=2]()
        %3 : int = prim::Constant[value=3]()
        %in_param : int[] = prim::ListConstruct(%0, %2, %3, %1)
        %a_perm : Tensor = aten::permute(%a_quant, %in_param)
        %w_perm : Tensor = aten::permute(%w_quant, %in_param)
        %w_packed = quantized::conv_prepack(%w_perm, %stride, %padding, %dilation, %groups)
        %r = quantized::conv2d(%a_perm, %w_packed, %b_quant, %stride, %padding, %dilation, %groups, %r_scale, %r_zero_point)
        %out_param : int[] = prim::ListConstruct(%0, %3, %1, %2)
        %r_perm = aten::permute(%r, %out_param)
        return (%r_perm))";
  SubgraphRewriter rewriter;
  rewriter.RegisterRewritePattern(pattern, replacement);
  rewriter.runOnGraph(graph);
}

struct ConvBNParameters {
  at::Tensor conv_w;
  at::Tensor conv_b;
  at::Tensor bn_rm;
  at::Tensor bn_rv;
  double bn_eps = 0.0;
  at::Tensor bn_w;
  at::Tensor bn_b;
};

/**
 * Given the current weight and bias tensors of a Conv2d module and parameters
 * of the BatchNorm2d module we're folding with, compute the updated values for
 * the weight and bias.
 *
 * The function is basically copied from torch/nn/utils/fusion.py
 */
static std::tuple<at::Tensor, at::Tensor> computeUpdatedConvWeightAndBias(
    const ConvBNParameters& p) {
  at::Tensor bn_var_rsqrt = at::rsqrt(p.bn_rv + p.bn_eps);
  at::Tensor new_w = p.conv_w * (p.bn_w * bn_var_rsqrt).reshape({-1, 1, 1, 1});
  at::Tensor new_b = (p.conv_b - p.bn_rm) * bn_var_rsqrt * p.bn_w + p.bn_b;
  return std::make_tuple(new_w, new_b);
}

static bool tryExtractingConvBNParameters(
    script::Module& conv,
    script::Module& bn,
    ConvBNParameters& r) {
  if (!conv.find_parameter("weight") || !bn.find_parameter("weight") ||
      !bn.find_parameter("bias")) {
    return false;
  }
  if (!bn.find_attribute("running_mean") || !bn.find_attribute("running_var") ||
      !bn.get_attribute("running_mean").isTensor() ||
      !bn.get_attribute("running_var").isTensor()) {
    return false;
  }

  r.bn_rm = bn.get_attribute("running_mean").toTensor();
  r.bn_rv = bn.get_attribute("running_var").toTensor();
  r.bn_eps = 1e-5; // TODO: allow access to the actual value. NOLINT
                   // Now we cannot do it because we inline all fields that are
                   // in __constants__ and lose all tracks of them.
  r.bn_w = bn.get_parameter("weight");
  r.bn_b = bn.get_parameter("bias");

  r.conv_w = conv.get_parameter("weight");
  if (conv.find_parameter("bias")) {
    r.conv_b = conv.get_parameter("bias");
  } else {
    r.conv_b = at::zeros_like(r.bn_rm);
  }

  return true;
}

void FoldConvBatchNorm2d(const script::Module& module) {
  std::string pattern = R"IR(
graph(%self, %x):
    %conv_submodule = match::module[name="Conv2d"](%self)
    %conv_out = prim::CallMethod[name="forward"](%conv_submodule, %x)
    %bn_submodule = match::module[name="BatchNorm2d"](%self)
    %bn_out = prim::CallMethod[name="forward"](%bn_submodule, %conv_out)
    return (%bn_out))IR";

  Graph pattern_graph;
  std::unordered_map<std::string, Value*> vmap;
  script::parseIR(pattern, &pattern_graph, vmap);
  Value* pattern_conv_out = vmap["conv_out"];
  Value* pattern_bn_out = vmap["bn_out"];
  Value* pattern_conv_submodule = vmap["conv_submodule"];
  Value* pattern_bn_submodule = vmap["bn_submodule"];
  Node* pattern_conv = pattern_conv_out->node();
  Node* pattern_bn = pattern_bn_out->node();

  // We will put submodules into this worklist and keep processing items from it
  // one by one. We start by just putting the top module there.
  std::stack<script::Module> worklist({module});
  while (!worklist.empty()) {
    script::Module current = worklist.top();
    worklist.pop();

    // Queue submodules for processing
    for (const script::Module& submodule : current.get_modules()) {
      worklist.push(submodule);
    }

    // Process forward method of the current module
    std::unordered_map<Value*, Value*> rewrite_map;
    std::vector<Value*> values_to_rewrite;
    std::unordered_set<Node*> nodes_to_delete;

    script::Method method = current.get_method("forward");
    GRAPH_DUMP(
        current.name().name() + "::forward() before Conv2d-BatchNorm2d folding",
        method.graph());
    const auto& matches = findPatternMatches(pattern_graph, *method.graph());

    for (const Match& match : matches) {
      GRAPH_DEBUG("Checking next match...");
      Node* matched_conv = match.nodes_map.at(pattern_conv);
      Node* matched_bn = match.nodes_map.at(pattern_bn);
      Node* matched_conv_submodule =
          match.values_map.at(pattern_conv_submodule)->node();
      Node* matched_bn_submodule =
          match.values_map.at(pattern_bn_submodule)->node();

      TORCH_INTERNAL_ASSERT(matched_conv_submodule->kind() == prim::GetAttr);
      TORCH_INTERNAL_ASSERT(matched_bn_submodule->kind() == prim::GetAttr);

      script::Module conv_submodule =
          current.get_module(matched_conv_submodule->s(Symbol::attr("name")));
      script::Module bn_submodule =
          current.get_module(matched_bn_submodule->s(Symbol::attr("name")));

      ConvBNParameters params;
      if (!tryExtractingConvBNParameters(
              conv_submodule, bn_submodule, params)) {
        GRAPH_DEBUG(
            "Conv and BN modules didn't have all required parameters or attributes...");
        continue;
      }

      // We are using a separate vector for saving Values we want to rewrite to
      // make sure that the order in which we perform these transformations is
      // deterministic. Iterating through keys of rewrite_map would result in
      // non-determinism that might not manifest as a bug now, but can bite us
      // later.
      values_to_rewrite.push_back(matched_bn->output());
      rewrite_map[matched_bn->output()] = matched_conv->output();
      GRAPH_UPDATE(
          "Rewriting %",
          matched_bn->output()->debugName(),
          " with %",
          matched_conv->output()->debugName());

      nodes_to_delete.insert(matched_bn);
      GRAPH_UPDATE("Deleting ", *matched_bn);

      auto new_w_b = computeUpdatedConvWeightAndBias(params);
      params.conv_w.set_data(std::get<0>(new_w_b));
      params.conv_b.set_data(std::get<1>(new_w_b));
    }

    // Perform planned rewritings
    for (auto v : values_to_rewrite) {
      v->replaceAllUsesWith(rewrite_map.at(v));
    }

    // Perform planned deletions
    for (auto n : nodes_to_delete) {
      n->removeAllInputs();
    }
    for (auto n : nodes_to_delete) {
      n->destroy();
    }
  }
}

} // namespace jit
} // namespace torch