Label Pattern Partitions

include/tvm/relay/function.h

@@ -141,6 +141,8 @@ constexpr const char* kSkipOptimization = "SkipOptimization";
 constexpr const char* kComposite = "Composite";
 /*! \brief Mark the function to be inlined. */
 constexpr const char* kInline = "Inline";
+/*! \brief Indicate the function was created by the Pattern Partitioning Pass. */
+constexpr const char* kPartitionedFromPattern = "PartitionedFromPattern";
 }  // namespace attr
 
 }  // namespace relay

python/tvm/relay/dataflow_pattern/__init__.py

@@ -112,21 +112,23 @@ def match(self, expr: Expr) -> bool:
         """
         return match(self, expr)
 
-    def partition(self, expr: Expr) -> bool:
+    def partition(self, expr: Expr, attrs=None) -> Expr:
         """
         Parition the expression into functions defined by this pattern
 
         Parameters
         ----------
         expr : tvm.relay.Expr
             The expression to match.
+        attrs : Optional[Dict[str, Object]]
+            A dictionary of Attribute name/values to add to the paritioned function
 
         Returns
         -------
         result : tvm.relay.Expr
             The Expression with matched subgraphs replaced by function calls to that subgraph
         """
-        return partition(self, expr)
+        return partition(self, expr, attrs)
 
     def dominates(self, parent, path=None):
         """
@@ -562,7 +564,7 @@ def rewrite(callbacks, expr: Expr) -> Expr:
 
     return ffi.rewrite(tmp, expr)
 
-def partition(pattern: DFPattern, expr: Expr) -> Expr:
+def partition(pattern: DFPattern, expr: Expr, attrs=None) -> Expr:
     """
     Parition the expression into a series of functions that match the pattern
 
@@ -572,10 +574,12 @@ def partition(pattern: DFPattern, expr: Expr) -> Expr:
         The pattern to match
     expr : tvm.relay.Expr
         The expression to split into functions
+    expr : Optional[Dict[str, Object]]
+        A dict of attributes to apply to the partitioned function
 
     Returns
     -------
     result : tvm.relay.Expr
         The Expression with matched subgraphs replaced by function calls to that subgraph
     """
-    return ffi.partition(pattern, expr)
+    return ffi.partition(pattern, expr, attrs)

src/relay/ir/dataflow_matcher.cc

@@ -386,6 +386,7 @@ class PatternGrouper : protected MixedModeVisitor {
     Expr root_node;
     int gid;
     Map<DFPattern, Array<Expr>> matched_nodes;
+    std::string name;
     Function function;
     Array<Expr> args;
   };
@@ -409,19 +410,25 @@ class PatternGrouper : protected MixedModeVisitor {
   }
 
  protected:
+  using ExprVisitor::VisitExpr_;
   void VisitLeaf(const Expr& pre) override {
     if (matcher_->Match(pattern_, pre)) {
       CreateGroup(pre);
     }
   }
-
+  void VisitExpr_(const FunctionNode* op) override {
+    if (op->attrs->dict.count(attr::kPartitionedFromPattern) == 0) {
+      ExprVisitor::VisitExpr_(op);
+    }
+  }
   /* \brief Creates a new set of nodes based on Group inputs, used to create functions and perform
    * group overlap analysis */
   class MatchExtractor : public ExprMutator {
    public:
     explicit MatchExtractor(const std::unordered_map<Expr, Var, ObjectHash, ObjectEqual>& inputs)
         : inputs_(inputs) {}
     const std::unordered_map<Expr, Expr, ObjectHash, ObjectEqual>& GetMemo() { return this->memo_; }
+    const std::string& GetName() { return name_; }
 
    protected:
     Expr VisitExpr(const Expr& pre) override {
@@ -430,6 +437,46 @@ class PatternGrouper : protected MixedModeVisitor {
       }
       return ExprMutator::VisitExpr(pre);
     }
+    Expr VisitExpr_(const TupleNode* op) override {
+      auto out = ExprMutator::VisitExpr_(op);
+      name_ += "Tuple_";
+      return out;
+    };
+    Expr VisitExpr_(const FunctionNode* op) override {
+      auto out = ExprMutator::VisitExpr_(op);
+      name_ += "Function";
+      return out;
+    };
+    Expr VisitExpr_(const CallNode* call_node) override {
+      auto out = ExprMutator::VisitExpr_(call_node);
+      if (auto operation = call_node->op.as<OpNode>()) {
+        name_ += operation->name + "_";
+      } else {
+        name_ += "Call_";
+      }
+      return out;
+    };
+    Expr VisitExpr_(const LetNode* op) override {
+      auto out = ExprMutator::VisitExpr_(op);
+      name_ += "Let_";
+      return out;
+    };
+    Expr VisitExpr_(const IfNode* op) override {
+      auto out = ExprMutator::VisitExpr_(op);
+      name_ += "If_";
+      return out;
+    };
+    Expr VisitExpr_(const TupleGetItemNode* op) override {
+      auto out = ExprMutator::VisitExpr_(op);
+      name_ += "TupleGetItem" + std::to_string(op->index) + "_";
+      return out;
+    };
+    Expr VisitExpr_(const MatchNode* op) override {
+      auto out = ExprMutator::VisitExpr_(op);
+      name_ += "Match_";
+      return out;
+    };
+    std::string name_;
     const std::unordered_map<Expr, Var, ObjectHash, ObjectEqual> inputs_;
   };
 
@@ -487,7 +534,7 @@ class PatternGrouper : protected MixedModeVisitor {
     // Verify the pattern still holds
     CHECK(DFPatternMatcher(body).Match(pattern_, body));
     group.function = Function(params, body, NullValue<Type>(), Array<TypeVar>());
-
+    group.name = extractor.GetName();
     // Check to make sure we aren't overlapping with another group
     // The MatchExtractor will create a new graph by replacing nodes that match the inputs of the
     // pattern with the input FunctionVar* Variables. The resulting memoization map will only
@@ -612,10 +659,12 @@ TVM_REGISTER_GLOBAL("relay.dataflow_pattern.rewrite").set_body_typed(RewritePatt
  */
 class PatternPartitioner : protected MixedModeMutator {
  public:
-  Expr Partition(const DFPattern& pattern, const Expr& pre) {
+  Expr Partition(const DFPattern& pattern, const Expr& pre,
+                 const Map<std::string, ObjectRef>& attrs) {
     auto grouper = PatternGrouper();
     groups_ = grouper.GroupMatches(pattern, pre);
     gid_assignments_ = grouper.GetGIDAssignments();
+    attrs_ = attrs;
     return this->VisitExpr(pre);
   }
 
@@ -625,7 +674,13 @@ class PatternPartitioner : protected MixedModeMutator {
     for (size_t i = 0; i < group.args.size(); ++i) {
       args.push_back(memo_[group.args[i]]);
     }
-    return Call(group.function, args);
+    Function func = WithAttr(group.function, attr::kPartitionedFromPattern, String(group.name));
+    if (!attrs_.empty()) {
+      for (auto kv : attrs_) {
+        func = WithAttr(std::move(func), kv.first, kv.second);
+      }
+    }
+    return Call(func, args);
   }
 
   Expr DispatchVisitExpr(const Expr& pre) override {
@@ -636,15 +691,19 @@ class PatternPartitioner : protected MixedModeMutator {
     return post;
   }
 
+  Map<std::string, ObjectRef> attrs_;
   std::vector<PatternGrouper::Group> groups_;
   std::unordered_map<Expr, int, ObjectHash, ObjectEqual> gid_assignments_;
 };
 
-Expr PartitionPattern(DFPattern pattern, Expr expr) {
-  return PatternPartitioner().Partition(pattern, expr);
+Expr PartitionPattern(DFPattern pattern, Expr expr, Map<std::string, ObjectRef> attrs) {
+  return PatternPartitioner().Partition(pattern, expr, attrs);
 }
 
-TVM_REGISTER_GLOBAL("relay.dataflow_pattern.partition").set_body_typed(PartitionPattern);
+TVM_REGISTER_GLOBAL("relay.dataflow_pattern.partition")
+    .set_body_typed([](DFPattern pattern, Expr expr, Map<std::string, ObjectRef> attrs) {
+      return PartitionPattern(pattern, expr, attrs);
+    });
 
 }  // namespace relay
 }  // namespace tvm

tests/python/relay/test_dataflow_pattern.py

@@ -702,6 +702,41 @@ def test_algebraic_simplify():
 
     assert tvm.ir.structural_equal(algebraic_simplify((x + zero * y) / one + (y * one) - zero / x), x + y)
 
+def test_double_partition():
+    # Pattern 1
+    conv2d_p = is_op('nn.conv2d')(wildcard(), wildcard())
+    bias_add_p = is_op("nn.bias_add")(conv2d_p, wildcard())
+    relu_p = is_op('nn.relu')(bias_add_p)
+
+    # Graph
+    x = relay.var('input')
+    w = relay.var('weight')
+    b = relay.var('bias')
+    w2 = relay.var('weight')
+    b2 = relay.var('bias')
+    conv2d = relay.op.nn.conv2d(x, w)
+    bias_add = relay.op.nn.bias_add(conv2d, b)
+    relu = relay.op.nn.relu(bias_add)
+    conv2d2 = relay.op.nn.conv2d(relu, w2)
+    bias_add2 = relay.op.nn.bias_add(conv2d2, b2)
+
+    partitioned = bias_add2
+    for pat, label in [(relu_p, "conv_bias_relu"), (bias_add_p, "conv_bias")]:
+        partitioned = pat.partition(partitioned, {"Composite": label})
+
+
+    inpf = relay.var("input")
+    weightf = relay.var("weight")
+    biasf = relay.var("bias")
+    func0 = relay.Function([inpf, weightf, biasf], relay.op.nn.relu(relay.op.nn.bias_add(relay.op.nn.conv2d(inpf, weightf), biasf))).with_attr("Composite", "conv_bias_relu").with_attr("PartitionedFromPattern","nn.conv2d_nn.bias_add_nn.relu_")
+    inpf = relay.var("input")
+    weightf = relay.var("weight")
+    biasf = relay.var("bias")
+    func1 = relay.Function([inpf, weightf, biasf], relay.op.nn.bias_add(relay.op.nn.conv2d(inpf, weightf), biasf)).with_attr("Composite", "conv_bias").with_attr("PartitionedFromPattern","nn.conv2d_nn.bias_add_")
+
+    expected = func1(func0(x, w, b), w2, b2)
+    assert tvm.ir.structural_equal(partitioned, expected)
+
 def test_partition_dominator():
     # Pattern
     is_conv2d = is_op('nn.conv2d')(wildcard(), wildcard())
@@ -721,10 +756,10 @@ def generate_diamond(inp, weight):
     out = generate_diamond(inp*inp, weight*weight)
     # Check
     partitioned = diamond.partition(out)
-
+    
     i = relay.Var("input")
     w = relay.Var("weight")
-    f = relay.Function([i, w], generate_diamond(i, w))
+    f = relay.Function([i, w], generate_diamond(i, w)).with_attr("PartitionedFromPattern","nn.conv2d_nn.relu_nn.relu_nn.leaky_relu_add_")
     assert tvm.ir.structural_equal(partitioned, f(inp*inp, weight*weight))
 
 def test_quadruple_partition_dominator():
@@ -783,10 +818,16 @@ def nested_diamond(inp, weight):
                 )
 
     functions = []
-    for f in [classic_diamond, deeper_diamond, single_branch, nested_diamond]:
+    partition_names = [
+        "nn.conv2d_nn.relu_nn.relu_nn.leaky_relu_add_",
+        "nn.conv2d_nn.relu_nn.relu_tanh_nn.leaky_relu_add_",
+        "nn.conv2d_nn.relu_nn.relu_tanh_add_",
+        "nn.conv2d_nn.relu_add_tanh_nn.leaky_relu_add_"
+    ]
+    for i, f in enumerate([classic_diamond, deeper_diamond, single_branch, nested_diamond]):
         inpf = relay.var("input")
         weightf = relay.var("weight")
-        functions.append(relay.Function([inpf, weightf], f(inpf, weightf)))
+        functions.append(relay.Function([inpf, weightf], f(inpf, weightf)).with_attr("PartitionedFromPattern", partition_names[i]))
 
     reference = functions[3](
                     functions[2](
@@ -816,7 +857,7 @@ def test_parition_batchnorm():
     betaf = relay.var('betaf')
     gammaf = relay.var('gammaf')
     # Put the arguments in toplogological order for the reference
-    f = relay.Function([gammaf, xf, meanf, varf, betaf], get_BN(xf, varf, meanf, betaf, gammaf))
+    f = relay.Function([gammaf, xf, meanf, varf, betaf], get_BN(xf, varf, meanf, betaf, gammaf)).with_attr("PartitionedFromPattern","subtract_multiply_add_sqrt_divide_add_")
 
     partitioned = BatchnormCallback().pattern.partition(BN)
     assert tvm.ir.structural_equal(partitioned, f(gamma, x, mean, var, beta))
@@ -836,14 +877,14 @@ def test_parition_double_batchnorm():
     meanf = relay.var('meanf')
     betaf = relay.var('betaf')
     gammaf = relay.var('gammaf')
-    f1 = relay.Function([gammaf, xf, meanf, varf, betaf], get_BN(xf, varf, meanf, betaf, gammaf))
+    f1 = relay.Function([gammaf, xf, meanf, varf, betaf], get_BN(xf, varf, meanf, betaf, gammaf)).with_attr("PartitionedFromPattern","subtract_multiply_add_sqrt_divide_add_")
     # The paritioner doesn't replace duplicates, so we use two copies of the function
     xf2 = relay.var('xf2')
     varf2 = relay.var('varf2')
     meanf2 = relay.var('meanf2')
     betaf2 = relay.var('betaf2')
     gammaf2 = relay.var('gammaf2')
-    f2 = relay.Function([gammaf2, xf2, meanf2, varf2, betaf2], get_BN(xf2, varf2, meanf2, betaf2, gammaf2))
+    f2 = relay.Function([gammaf2, xf2, meanf2, varf2, betaf2], get_BN(xf2, varf2, meanf2, betaf2, gammaf2)).with_attr("PartitionedFromPattern","subtract_multiply_add_sqrt_divide_add_")
 
     partitioned = BatchnormCallback().pattern.partition(BN2)
     reference = f2(gamma, f1(gamma, x, mean, var, beta), mean, var, beta)