junrushao1994 commented on a change in pull request #5962: URL: https://github.com/apache/incubator-tvm/pull/5962#discussion_r449966709
########## File path: src/ansor/compute_dag.cc ########## @@ -0,0 +1,505 @@ +/* + * 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. +// Results are stored in ops +void TopoSortOps(const Array<te::Tensor>& tensors, Array<te::Operation>* ops) { + 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 + ops->clear(); + + 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])); + } + } + } +} + +// 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}; +}; + +ComputeDAG::ComputeDAG(Array<te::Tensor> tensors) { + auto node = make_object<ComputeDAGNode>(); + FlopEstimator estimator; + Array<te::Operation> ops; + node->tensors = std::move(tensors); + TopoSortOps(node->tensors, &ops); + node->ops = std::move(ops); + node->flop_ct = estimator.EstimateFlop(node->ops); + node->init_state = State(node->ops); + data_ = std::move(node); +} + +// Update the te::stage to tir::IterVar axis mapping +void UpdateStageAxis(const te::Stage& stage, StageToAxesMap* stage_to_axes) { + if (auto pop = stage->op.as<te::ComputeOpNode>()) { + Array<IterVar> axes; + for (const auto& axis : pop->axis) { + axes.push_back(axis); + } + for (const auto& axis : pop->reduce_axis) { + axes.push_back(axis); + } + stage_to_axes->Set(stage, std::move(axes)); + } else if (stage->op->IsInstance<te::PlaceholderOpNode>()) { + {} // do nothing on Placeholder + } else { + LOG(FATAL) << "Invalid op " << stage->op; + } +} + +std::pair<te::Schedule, Array<te::Tensor>> ComputeDAG::ApplySteps( + const Array<Step>& transform_steps, Array<te::Stage>* stages, + StageToAxesMap* stage_to_axes) const { + // Temporal object to be used if the input pointer is nullptr + Array<te::Stage> temp_stages; + StageToAxesMap temp_stage_to_axes; + if (stages == nullptr) { + stages = &temp_stages; + } + if (stage_to_axes == nullptr) { + stage_to_axes = &temp_stage_to_axes; + } + Array<te::Operation> ops; + for (const auto& op : operator->()->ops) { + if (!op->IsInstance<te::PlaceholderOpNode>()) { + ops.push_back(op); + } + } + // Create the initial schedule + te::Schedule schedule = te::create_schedule({ops.back()}); + + // init axes + for (const auto& x : operator->()->ops) { + const te::Stage& stage = schedule.operator[](x); + stages->push_back(stage); + UpdateStageAxis(stage, stage_to_axes); + } + + // Use complete rate for the study in the paper + const char* complete_rate_str = getenv("ANSOR_PROGRAM_COMPLETE_RATE"); + double complete_rate = -1.0; + if (complete_rate_str) { + complete_rate = std::stod(complete_rate_str); + } + size_t ct = 0; + // Apply the history steps to TVM schedule + for (const auto& step : transform_steps) { + if (complete_rate >= 0 && ct++ > transform_steps.size() * complete_rate) { + break; + } + // Call each step's ApplyToSchedule method + // Note: some steps have extra parameters that must be passed and they may need different + // return value, so the ApplyToSchedule is not able to be merged to single interface + if (auto ps = step.as<ReorderStepNode>()) { + ps->ApplyToSchedule(stages, stage_to_axes); + } else if (auto ps = step.as<SplitStepNode>()) { + ps->ApplyToSchedule(stages, stage_to_axes); + } else if (auto ps = step.as<FuseStepNode>()) { + ps->ApplyToSchedule(stages, stage_to_axes); + } else { + LOG(FATAL) << "Invalid Step"; + } + } + + return std::make_pair(schedule, operator->()->tensors); +} + +String ComputeDAG::PrintStepsAsPython(const Array<Step>& transform_steps) const { + Array<te::Stage> stages; + StageToAxesMap stage_to_axes; + Array<te::Operation> ops; + for (const auto& op : operator->()->ops) { + if (!op->IsInstance<te::PlaceholderOpNode>()) { + ops.push_back(op); + } + } + // Create the initial schedule + te::Schedule schedule = te::create_schedule({ops.back()}); Review comment: Hmmm let’s make it correct here ---------------------------------------------------------------- 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: us...@infra.apache.org
