[Collage] CombinerRule and CandidatePartition::EstimateCost (#12078)

* [Collage] CombinerRule and CandidatePartition::EstimateCost

See https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md.

We complete the PartitionRule sub-class hierarchy with the addition of
CombinePartitionRule, which allows disjoint candidate partitions to be
unioned based on simple rules.
 - By TOpPattern kind, eg a kOutElemwiseFusable and kBroadcast.
 - A tuple argument with injective fields.
 - The projection from an injective group (obviously of tuple type)
 - Combinations of the above.
These let us mimic many common fusion strategies, including TVMs, so that
the candidates explored during Collage search are as large as possible to
expose possible fusion opportunities but no larger.

Also completes CandidatePartition with the EstimateCost method, which is
used during search to construct a stand-alone IRModule for latency estimation.

Finish units tests for PartitionRule and CandidatePartition.

* - fix relay.collage ffi prefix.

src/relay/collage/candidate_function_cache.cc

@@ -0,0 +1,49 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/collage/candidate_function_cache.cc
+ * \brief A cache of the unique global name and costs for partitioned functions.
+ */
+
+#include "./candidate_function_cache.h"
+
+namespace tvm {
+namespace relay {
+namespace collage {
+
+CandidateFunctionCache::Entry& CandidateFunctionCache::GetEntry(const std::string& label,
+                                                                const Function& function) {
+  auto itr = cache_.find(function);
+  if (itr == cache_.end()) {
+    String compiler = function->GetAttr<String>(attr::kCompiler, String("tvm")).value();
+    std::string global_symbol_name = name_supply_->Fresh({compiler, label});
+    GlobalVar global_symbol(std::move(global_symbol_name), function->checked_type());
+    itr = cache_.emplace(function, Entry(std::move(global_symbol))).first;
+  }
+  return itr->second;
+}
+
+GlobalVar CandidateFunctionCache::GetGlobalSymbol(const Function& function) {
+  return GetEntry(/*label=*/"", function).global_symbol;
+}
+
+}  // namespace collage
+}  // namespace relay
+}  // namespace tvm

src/relay/collage/candidate_function_cache.h

@@ -0,0 +1,79 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/collage/candidate_function_cache.h
+ * \brief A cache of the unique global symbol name and cost for partitioned functions.
+ */
+
+#ifndef TVM_RELAY_COLLAGE_CANDIDATE_FUNCTION_CACHE_H_
+#define TVM_RELAY_COLLAGE_CANDIDATE_FUNCTION_CACHE_H_
+
+#include <tvm/relay/function.h>
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+
+#include "../transforms/compiler_function_utils.h"
+#include "./cost.h"
+#include "./name_supply.h"
+
+namespace tvm {
+namespace relay {
+namespace collage {
+
+/*!
+ * \brief A cache of the unique global symbol and cost for functions extracted to represent
+ * partitions. If two functions are structurally equal (which includes equality of their "Compiler"
+ * attributes) then they will share the same global symbol and estimated cost. We rely on the
+ * function's attributes to distinguish partitions which are structurally the same graph but
+ * intended for different targets.
+ */
+class CandidateFunctionCache : public transform::GlobalSymbolCache {
+ public:
+  explicit CandidateFunctionCache(std::shared_ptr<NameSupply> name_supply)
+      : name_supply_(std::move(name_supply)) {}
+
+  struct Entry {
+    GlobalVar global_symbol;
+    Cost cost = Cost::Unknown();  // Filled in when have estimated cost.
+
+    explicit Entry(GlobalVar global_symbol) : global_symbol(std::move(global_symbol)) {}
+  };
+
+  /*!
+   * \brief Returns the unique entry for \p function. If no such entry already exists, create it
+   * and assign it a unique global symbol name.
+   */
+  Entry& GetEntry(const std::string& label, const Function& function);
+
+  GlobalVar GetGlobalSymbol(const Function& function) final;
+
+ private:
+  std::shared_ptr<NameSupply> name_supply_;
+  std::unordered_map<Function, Entry, StructuralHash, StructuralEqual> cache_;
+};
+
+}  // namespace collage
+}  // namespace relay
+}  // namespace tvm
+
+#endif  // TVM_RELAY_COLLAGE_CANDIDATE_FUNCTION_CACHE_H_

src/relay/collage/candidate_partition.cc

@@ -24,8 +24,12 @@
 
 #include "./candidate_partition.h"
 
+#include <tvm/relay/analysis.h>
 #include <tvm/relay/attrs/memory.h>
+#include <tvm/relay/transform.h>
 
+#include "../transforms/compiler_function_utils.h"
+#include "./candidate_function_cache.h"
 #include "./candidate_set.h"
 #include "./partition_rule.h"
 #include "./partition_spec.h"
@@ -106,6 +110,102 @@ std::string CandidatePartitionNode::ToString() const {
   return os.str();
 }
 
+namespace {
+/*!
+ * \brief If function's body is a call to an inlined "Primitive" function, return it.
+ * Otherwise return function directly.
+ */
+Function GetPrimitiveFunction(const Function& function) {
+  if (const auto* call_node = function->body.as<CallNode>()) {
+    if (const auto* function_node = call_node->op.as<FunctionNode>()) {
+      if (function_node->HasNonzeroAttr(attr::kPrimitive)) {
+        return GetRef<Function>(function_node);
+      }
+    }
+  }
+  return function;
+}
+
+/*!
+ * \brief Eta-expand any tuple arguments of \p function. Ie rewrite:
+ * \code
+ *   f(x: (t1, t2)) { ... x ... }
+ * \endcode
+ * to
+ * \code
+ *   f(x_1: t1, x_2: t2) { ... (x_1, x_2) ... }
+ * \endcode
+ */
+Function EtaExpandTuples(const Function& function) {
+  Map<Var, Expr> subst;
+  Array<Var> new_params;
+  for (const auto& param : function->params) {
+    std::vector<TensorType> tensor_types = FlattenTupleType(param->type_annotation);
+    if (tensor_types.size() == 1) {
+      new_params.push_back(param);
+    } else {
+      Array<Expr> fields;
+      for (size_t i = 0; i < tensor_types.size(); ++i) {
+        Var new_param(param->name_hint() + "_" + std::to_string(i), tensor_types[i], param->span);
+        new_param->checked_type_ = tensor_types[i];
+        new_params.push_back(new_param);
+        fields.push_back(new_param);
+      }
+      Tuple new_tuple(fields);
+      subst.Set(param, new_tuple);
+    }
+  }
+  if (subst.empty()) {
+    return function;
+  }
+  return WithFields(function, new_params, Bind(function->body, subst));
+}
+
+}  // namespace
+
+Cost CandidatePartitionNode::EstimatedCost(
+    const DataflowGraph& dataflow_graph, const CostEstimator& cost_estimator,
+    const std::shared_ptr<CandidateFunctionCache>& cache) const {
+  if (cost_.is_unknown()) {
+    VLOG_CONTEXT << "spec " << partition_spec_name();
+    Function extracted_function = sub_graph_->ExtractAsFunction(dataflow_graph);
+    VLOG(2) << "Extracted function:" << std::endl << PrettyPrint(extracted_function);
+    extracted_function = EtaExpandTuples(extracted_function);
+    VLOG(2) << "Validating function:" << std::endl << PrettyPrint(extracted_function);
+    String error = partition_spec()->validate_sub_graph_func_(extracted_function);
+    if (!error.empty()) {
+      cost_ = Cost::Invalid();
+      VLOG(1) << "Unable to rewrite function: " << error;
+    } else {
+      // The extracted function may be the eta-expansion of a "Primitive" function.
+      // If so we want the cached external name and cost to be w.r.t. that function
+      // rather than the outer so that we'll get a cache hit when we outline functions
+      // in the final program.
+      Function primitive_function = GetPrimitiveFunction(extracted_function);
+      CandidateFunctionCache::Entry& entry =
+          cache->GetEntry(sub_graph_->label_, primitive_function);
+      if (entry.cost.is_unknown()) {
+        IRModule mod = IRModule::FromExpr(extracted_function);
+        VLOG(1) << "Outlining:" << std::endl << PrettyPrint(mod);
+        mod = OutlineCompilerFunctions(cache)(mod);
+        VLOG(1) << "Estimating cost of:" << std::endl
+                << PrettyPrint(mod) << std::endl
+                << "using target " << target()->ToDebugString();
+        entry.cost = cost_estimator->Estimate(mod, target(),
+                                              /*needs_tvm_tuning=*/!target().IsExternalCodegen());
+        VLOG(1) << "Measured cost as " << entry.cost.ToString();
+      } else {
+        VLOG(1) << "Reusing cost " << entry.cost.ToString()
+                << " cached in candidate function cache";
+      }
+      cost_ = entry.cost;
+    }
+  } else {
+    VLOG(1) << "Reusing cost " << cost_.ToString() << " cached in candidate";
+  }
+  return cost_;
+}
+
 CandidatePartition::CandidatePartition(String rule_name, SubGraph sub_graph,
                                        ObjectRef /* actually PartitionSpec */ spec, Cost cost) {
   auto node = runtime::make_object<CandidatePartitionNode>();

src/relay/collage/candidate_partition.h

@@ -32,7 +32,10 @@
 #include <string>
 #include <vector>
 
+#include "./candidate_function_cache.h"
 #include "./cost.h"
+#include "./cost_estimator.h"
+#include "./name_supply.h"
 #include "./sub_graph.h"
 
 namespace tvm {
@@ -93,6 +96,13 @@ class CandidatePartitionNode : public Object {
    */
   Target target() const;
 
+  /*!
+   * \brief Return the estimated cost of the candidate partition, using \p cost_estimator and
+   * \p cache.
+   */
+  Cost EstimatedCost(const DataflowGraph& dataflow_graph, const CostEstimator& cost_estimator,
+                     const std::shared_ptr<CandidateFunctionCache>& cache) const;
+
   /*!
    * \brief Returns a brief description of candidate suitable for debugging output.
    */