[GitHub] [incubator-tvm] mbrookhart commented on a change in pull request #5962: [Ansor][AutoTVM v2.0] Part 0: Ansor minimum system for auto schedule generating

[GitBox <gi...@apache.org>]

mbrookhart commented on a change in pull request #5962:
URL: https://github.com/apache/incubator-tvm/pull/5962#discussion_r450528796



##########
File path: src/ansor/compute_dag.h
##########
@@ -0,0 +1,122 @@
+/*
+ * 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 ansor/compute_dag.h
+ * \brief The Ansor computational graph and related program analyses.
+ *
+ * We convert a compute declaration described by `tvm.compute` (could be a single operator or a
+ * subgraph) to a ComputeDAG. It keeps the input/output tensors of the compute declaration,
+ * a list of all operations in the DAG as well as static analysis results for the DAG (e.g. the
+ * total float operation count, consumer/producer relations of each operation stage, whether an
+ * operation stage should be tiled/compute inlined ...). These analyses can help the search policy
+ * to make decisions during search process.
+ * ComputeDAG is also responsible for the interaction between Ansor `LoopState` and TVM schedule
+ * (e.g. applying the `LoopState` transform steps to TVM schedule, providing `LoopState` with extra
+ * information got from TVM schedule ...).
+ */
+
+#ifndef TVM_ANSOR_COMPUTE_DAG_H_
+#define TVM_ANSOR_COMPUTE_DAG_H_
+
+#include <tvm/te/schedule.h>
+
+#include <utility>
+
+#include "loop_state.h"
+
+namespace tvm {
+namespace ansor {
+
+/*! \brief The Ansor computational graph and related program analyses. */
+class ComputeDAGNode : public Object {
+ public:
+  /*! \brief Input and output tensors. */
+  Array<te::Tensor> tensors;

Review comment:
       Why not separate input and output tensors? This feels fragile.

##########
File path: python/tvm/ansor/compute_dag.py
##########
@@ -0,0 +1,141 @@
+# 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.
+
+""" Computational graph and its analysis tools """
+
+import hashlib
+
+import tvm._ffi
+from tvm.runtime import Object
+from tvm.te import PlaceholderOp, ComputeOp
+
+from .loop_state import State, StateObject
+from .utils import get_const_tuple
+from .workload_registry import workload_key_to_tensors
+
+from . import _ffi_api
+
+
+@tvm._ffi.register_object("ansor.ComputeDAG")
+class ComputeDAG(Object):
+    """
+    Computation declaration graph.
+
+    Parameters
+    ----------
+    compute : Union[List[Tensor], str]
+        `Tensor`s or workload key for a compute declaration.
+    """
+    def __init__(self, compute):
+        if isinstance(compute, str):
+            compute = workload_key_to_tensors(compute)
+        elif isinstance(compute, list):
+            for item in compute:
+                if not isinstance(item, tvm.te.Tensor):
+                    raise ValueError("The input of ComputeDAG should be a list of Tensor")
+        else:
+            raise ValueError("Invalid compute: " + compute + ". Expect a string or list of Tensor")
+        self.__init_handle_by_constructor__(_ffi_api.ComputeDAG, compute)
+
+    def get_init_state(self):
+        """ Get init state of this ComputeDAG.
+
+        Returns
+        -------
+        state : State
+            The initial State without any transform steps.
+        """
+        return State(self.init_state, self)
+
+    def apply_steps_from_state(self, state):
+        """
+        Apply transform steps according to the history of a State.
+
+        Parameters
+        ----------
+        state : Union[State, StateObject]
+            The target state to be applied to TVM schedule.
+
+        Returns
+        -------
+            A `te.schedule` and the target `te.Tensor`s to be used in `tvm.lower` or `tvm.build`
+        """
+        state_obj = state if isinstance(state, StateObject) else state.state_object
+        return _ffi_api.ComputeDAGApplyStepsFromState(self, state_obj)
+
+    def print_python_code_from_state(self, state):
+        """
+        Print transform steps in the history of a State as TVM's python schedule primitive.
+
+        Parameters
+        ----------
+        state : Union[State, StateObject]
+            The target state to be applied to TVM schedule.
+
+        Returns
+        -------
+        str : Str
+            The Python schedule code.
+        """
+        state_obj = state if isinstance(state, StateObject) else state.state_object
+        return _ffi_api.ComputeDAGPrintPythonCodeFromState(self, state_obj)
+
+    def infer_bound_from_state(self, state):
+        """
+        Infer bound for a state using TVM schedule.
+
+        State api supports to define a split step with its split factor to be a blank placeholder,
+        so sometimes we may get a State will incomplete iterator extent information.
+        And another situation is after some steps (for exp. compute_at), it may be hard to track
+        the extent change of all iterators.
+
+        We perform infer bound using TVM schedule and fill the State with those informations. After
+        applying this methods, the State is guaranteed to have complete interator extent
+        information.
+
+        Parameters
+        ----------
+        state : Union[State, StateObject]
+            The target state to be applied to TVM schedule.
+
+        Returns
+        -------
+        state : State
+            The State with complete bound information.
+        """
+        state_obj = state if isinstance(state, StateObject) else state.state_object
+        return State(_ffi_api.ComputeDAGInferBoundFromState(self, state_obj), self)
+
+    def __hash__(self):
+        # TODO(...): Implement this more carefully and move this to c++ as a member function

Review comment:
       Assign a name to the TODO

##########
File path: src/ansor/compute_dag.cc
##########
@@ -0,0 +1,479 @@
+/*
+ * 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 ansor/compute_dag.cc
+ * \brief Compute declaration graph and its related analysis tools.
+ */
+
+#include "compute_dag.h"
+
+#include <tvm/runtime/registry.h>
+#include <tvm/te/operation.h>
+#include <tvm/te/schedule.h>
+#include <tvm/te/schedule_pass.h>
+#include <tvm/tir/stmt_functor.h>
+
+#include <algorithm>
+#include <queue>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "loop_state.h"
+#include "utils.h"
+
+namespace tvm {
+namespace ansor {
+
+using namespace tvm::tir;
+
+TVM_REGISTER_NODE_TYPE(ComputeDAGNode);
+
+// Topo-sort ops from tensors according to their read-write relations.
+Array<te::Operation> TopoSortOps(const Array<te::Tensor>& tensors) {
+  std::unordered_map<const te::OperationNode*, int> degree;
+  std::unordered_map<const te::OperationNode*, std::vector<const te::OperationNode*>> edge_set;
+  std::unordered_map<const te::OperationNode*, int> priority;
+  std::unordered_set<const te::OperationNode*> visited;
+
+  // traverse to build edge_set and count degree
+  std::vector<const te::OperationNode*> stack;
+  stack.reserve(tensors.size());
+  for (const auto& x : tensors) {
+    stack.push_back(x->op.operator->());
+  }
+
+  int ct = 0;
+  while (!stack.empty()) {
+    const te::OperationNode* op = stack.back();
+    stack.pop_back();
+    if (visited.count(op)) {
+      continue;
+    }
+
+    priority[op] = ct;
+    ct++;
+    visited.insert(op);
+
+    if (op->IsInstance<te::PlaceholderOpNode>()) {
+      degree[op] = 0;
+    } else if (auto cop = GetRef<te::Operation>(op).as<te::ComputeOpNode>()) {
+      const Array<te::Tensor>& input_tensors = cop->InputTensors();
+      degree[op] = input_tensors.size();
+      for (const auto& ten : input_tensors) {
+        edge_set[ten->op.operator->()].push_back(op);
+        stack.push_back(ten->op.operator->());
+      }
+    } else {
+      LOG(FATAL) << "Unsupported op " << GetRef<te::Operation>(op);
+    }
+  }
+
+  // topo sort
+  Array<te::Operation> ops;
+
+  using Item = std::pair<const te::OperationNode*, int>;
+  auto cmp = [](const Item& left, const Item& right) { return left.second < right.second; };
+  std::priority_queue<Item, std::vector<Item>, decltype(cmp)> queue(cmp);
+  for (const auto& iter : degree) {
+    if (iter.second == 0) {
+      queue.push(Item(iter.first, priority[iter.first]));
+    }
+  }
+
+  ops.reserve(degree.size());
+  while (!queue.empty()) {
+    Item item = queue.top();
+    queue.pop();
+    ops.push_back(GetRef<te::Operation>(item.first));
+    for (const auto& dst : edge_set[item.first]) {
+      degree[dst] -= 1;
+      if (degree[dst] == 0) {
+        queue.push(Item(dst, priority[dst]));
+      }
+    }
+  }
+
+  return ops;
+}
+
+// Estimate number of float operations in an expression
+class FlopEstimator : public ExprFunctor<double(const PrimExpr& n)> {

Review comment:
       I'm a little uncomfortable using recursion here, are we should we wont overflow?

##########
File path: src/ansor/compute_dag.cc
##########
@@ -0,0 +1,479 @@
+/*
+ * 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 ansor/compute_dag.cc
+ * \brief Compute declaration graph and its related analysis tools.
+ */
+
+#include "compute_dag.h"
+
+#include <tvm/runtime/registry.h>
+#include <tvm/te/operation.h>
+#include <tvm/te/schedule.h>
+#include <tvm/te/schedule_pass.h>
+#include <tvm/tir/stmt_functor.h>
+
+#include <algorithm>
+#include <queue>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "loop_state.h"
+#include "utils.h"
+
+namespace tvm {
+namespace ansor {
+
+using namespace tvm::tir;
+
+TVM_REGISTER_NODE_TYPE(ComputeDAGNode);
+
+// Topo-sort ops from tensors according to their read-write relations.
+Array<te::Operation> TopoSortOps(const Array<te::Tensor>& tensors) {
+  std::unordered_map<const te::OperationNode*, int> degree;
+  std::unordered_map<const te::OperationNode*, std::vector<const te::OperationNode*>> edge_set;
+  std::unordered_map<const te::OperationNode*, int> priority;
+  std::unordered_set<const te::OperationNode*> visited;
+
+  // traverse to build edge_set and count degree
+  std::vector<const te::OperationNode*> stack;
+  stack.reserve(tensors.size());
+  for (const auto& x : tensors) {
+    stack.push_back(x->op.operator->());
+  }
+
+  int ct = 0;
+  while (!stack.empty()) {
+    const te::OperationNode* op = stack.back();
+    stack.pop_back();
+    if (visited.count(op)) {
+      continue;
+    }
+
+    priority[op] = ct;
+    ct++;
+    visited.insert(op);
+
+    if (op->IsInstance<te::PlaceholderOpNode>()) {
+      degree[op] = 0;
+    } else if (auto cop = GetRef<te::Operation>(op).as<te::ComputeOpNode>()) {
+      const Array<te::Tensor>& input_tensors = cop->InputTensors();
+      degree[op] = input_tensors.size();
+      for (const auto& ten : input_tensors) {
+        edge_set[ten->op.operator->()].push_back(op);
+        stack.push_back(ten->op.operator->());
+      }
+    } else {
+      LOG(FATAL) << "Unsupported op " << GetRef<te::Operation>(op);
+    }
+  }
+
+  // topo sort
+  Array<te::Operation> ops;
+
+  using Item = std::pair<const te::OperationNode*, int>;
+  auto cmp = [](const Item& left, const Item& right) { return left.second < right.second; };
+  std::priority_queue<Item, std::vector<Item>, decltype(cmp)> queue(cmp);
+  for (const auto& iter : degree) {
+    if (iter.second == 0) {
+      queue.push(Item(iter.first, priority[iter.first]));
+    }
+  }
+
+  ops.reserve(degree.size());
+  while (!queue.empty()) {
+    Item item = queue.top();
+    queue.pop();
+    ops.push_back(GetRef<te::Operation>(item.first));
+    for (const auto& dst : edge_set[item.first]) {
+      degree[dst] -= 1;
+      if (degree[dst] == 0) {
+        queue.push(Item(dst, priority[dst]));
+      }
+    }
+  }
+
+  return ops;
+}
+
+// Estimate number of float operations in an expression
+class FlopEstimator : public ExprFunctor<double(const PrimExpr& n)> {
+ public:
+  double EstimateFlop(const Array<te::Operation>& ops) {
+    double ret = 0;
+    for (const auto& op : ops) {
+      if (auto pop = op.as<te::ComputeOpNode>()) {
+        double num_element = AxisLengthProd(pop->axis);
+        if (num_element == -1) {
+          fail = true;
+          break;
+        }
+        double op_per_element = 0;
+        for (const auto& x : pop->body) {
+          op_per_element += VisitExpr(x);
+        }
+        ret += num_element * op_per_element;
+      } else if (op->IsInstance<te::PlaceholderOpNode>()) {
+        {}  // do nothing
+      } else {
+        LOG(FATAL) << "Invalid op type " << op;
+      }
+    }
+
+    return fail ? -1 : ret;
+  }
+
+  double VisitExpr_(const ReduceNode* op) final {
+    uint64_t num_iter = 1;
+    for (const auto& x : op->axis) {
+      if (auto imm = x->dom->extent.as<IntImmNode>()) {
+        num_iter *= imm->value;
+      } else {
+        fail = true;
+        num_iter = -1;
+      }
+    }
+    double body_flop = 0;
+    for (size_t i = 0; i < op->combiner->result.size(); ++i) {
+      body_flop += VisitExpr(op->combiner->result[i]);
+      body_flop += VisitExpr(op->source[i]);
+    }
+    return num_iter * body_flop;
+  }
+
+  double VisitExpr_(const FloatImmNode* op) final { return 0.0; }
+  double VisitExpr_(const IntImmNode* op) final { return 0.0; }
+  double VisitExpr_(const ProducerLoadNode* op) final { return 0.0; }
+
+  double VisitExpr_(const CastNode* op) final { return VisitExpr(op->value); }
+  double VisitExpr_(const VarNode* op) final { return 0.0; }
+
+  double VisitExpr_(const SelectNode* op) final {
+    return VisitExpr(op->condition) +
+           std::max(VisitExpr(op->true_value), VisitExpr(op->false_value));
+  }
+
+#define VisitBinary(Node) \
+  double VisitExpr_(const Node* op) final { return 1.0 + VisitExpr(op->a) + VisitExpr(op->b); }
+#define VisitUnary(Node) \
+  double VisitExpr_(const Node* op) final { return 1.0 + VisitExpr(op->a); }
+
+  VisitBinary(AddNode);
+  VisitBinary(SubNode);
+  VisitBinary(MulNode);
+  VisitBinary(DivNode);
+  VisitBinary(ModNode);
+  VisitBinary(FloorDivNode);
+  VisitBinary(FloorModNode);
+  VisitBinary(MaxNode);
+  VisitBinary(MinNode);
+  VisitBinary(EQNode);
+  VisitBinary(NENode);
+  VisitBinary(LTNode);
+  VisitBinary(LENode);
+  VisitBinary(GTNode);
+  VisitBinary(GENode);
+  VisitBinary(AndNode);
+  VisitBinary(OrNode);
+  VisitUnary(NotNode);
+
+  double VisitExpr_(const CallNode* op) final {
+    double ret = 0.0;
+    for (const auto& x : op->args) {
+      ret += VisitExpr(x);
+    }
+    return ret;
+  }
+
+  double VisitExprDefault_(const Object* op) final {
+    fail = true;
+    return -1.0;
+  }
+
+  bool fail{false};

Review comment:
       call this fail_ per Google Style Guide, especially with using it inside so many member methods.




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org