mindspore2022/mindspore/ccsrc/parallel/auto_parallel/edge_costmodel.cc

298 lines
13 KiB
C++

/**
* Copyright 2019 Huawei Technologies Co., Ltd
*
* Licensed 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.
*/
#include "parallel/auto_parallel/edge_costmodel.h"
#include <algorithm>
#include <functional>
#include <iterator>
#include <utility>
#include "parallel/auto_parallel/costmodel.h"
#include "parallel/auto_parallel/graph_costmodel.h"
#include "parallel/tensor_layout/tensor_redistribution.h"
namespace mindspore {
namespace parallel {
Status Edge::InitEdgeCost() {
bool has_available_cost = false;
for (auto &swc : prev_op_->GetStrategyCost()) {
MS_EXCEPTION_IF_NULL(swc);
pre_op_output_.emplace_back(std::make_pair(swc->strategy_ptr, swc->outputs_ptr));
}
for (auto &swc : next_op_->GetStrategyCost()) {
MS_EXCEPTION_IF_NULL(swc);
next_op_input_.emplace_back(std::make_pair(swc->strategy_ptr, swc->inputs_ptr));
}
if (is_identity_edge) {
for (auto &target_output : pre_op_output_) {
auto target_output_lyt = target_output.second[prev_op_output_index_].tensor_layout();
auto target_output_str = target_output.first;
for (auto &target_input : next_op_input_) {
auto target_input_lyt = target_input.second[next_op_input_index_].tensor_layout();
auto target_input_str = target_input.first;
if (target_output_lyt == target_input_lyt) {
CostPtrKey ck = {target_output_str, target_input_str};
CostPtr cost = std::make_shared<Cost>(0.0, 0.0);
MS_EXCEPTION_IF_NULL(cost);
cost->communication_without_parameter_ = 0.0;
cost->communication_with_partial_para_ = 0.0;
CostPtrList cl;
cl.push_back(cost);
(void)cost_map_.emplace(std::make_pair(ck, cl));
has_available_cost = true;
}
}
}
} else {
for (auto &target_output : pre_op_output_) {
auto target_output_lyt = target_output.second[prev_op_output_index_].tensor_layout();
auto target_output_str = target_output.first;
auto type_length = prev_op_->GetOutputTypeLengths()[prev_op_output_index_];
auto type = prev_op_->outputs_type()[prev_op_output_index_];
for (auto &target_input : next_op_input_) {
auto target_input_lyt = target_input.second[next_op_input_index_].tensor_layout();
auto target_input_str = target_input.first;
CostPtr cost;
if (GetRedistributionCost(target_output_lyt, target_input_lyt, type_length, type, &cost) != SUCCESS) {
MS_LOG(EXCEPTION) << "Failure: redistribution cost calculation failed";
}
MS_EXCEPTION_IF_NULL(cost);
MS_LOG(DEBUG) << "The redistribution cost: computation_cost: " << cost->computation_cost_
<< ", communication_cost: " << cost->communication_cost_
<< ", communication_without_parameter_: " << cost->communication_without_parameter_
<< ", communication_with_partial_para_: " << cost->communication_with_partial_para_ << ".";
// refine communication cost calculation for practice
RefineForPracticalCost(cost, true);
CostPtrKey ck = {target_output_str, target_input_str};
CostPtrList cl;
cl.push_back(cost);
(void)cost_map_.emplace(std::make_pair(ck, cl));
has_available_cost = true;
}
}
}
if (!has_available_cost) {
if (FULLY_USE_DEVICES) {
MS_LOG(EXCEPTION) << "Generating cost for edge: " << edge_name_
<< " failed, it may be caused by setting 'fully_use_devices' true. Try to set "
"'fully_use_devices' false.";
} else if (ELEMENTWISE_OP_STRA_FOLLOW) {
MS_LOG(EXCEPTION) << "Generating cost for edge: " << edge_name_
<< " failed, it may be caused by setting 'elementwise_op_strategy_follow' true. "
"Try to set 'elementwise_op_strategy_follow' false.";
}
MS_LOG(EXCEPTION) << "Generating cost for edge: " << edge_name_ << " failed.";
}
return Status::SUCCESS;
}
Status Edge::GetRedistributionCost(const TensorLayout &prev_op_output_layout, const TensorLayout &next_op_input_layout,
size_t type_length, TypePtr type, CostPtr *cost) {
MS_EXCEPTION_IF_NULL(prev_op_);
MS_EXCEPTION_IF_NULL(cost);
RankList dev_list = prev_op_->global_device_list();
TensorRedistribution tensor_redistribution(false);
// Init TensorRedistribution
if (tensor_redistribution.Init(prev_op_output_layout, next_op_input_layout, dev_list) == FAILED) {
MS_LOG(EXCEPTION) << "Failure: tensor_redistribution init failed.";
}
if (tensor_redistribution.ComputeCost() == FAILED) {
MS_LOG(EXCEPTION) << "Failure: tensor_redistribution ComputeCost failed.";
}
double comm_cost = tensor_redistribution.comm_cost();
double forward_comm_cost = tensor_redistribution.forward_comm_cost();
double backward_comm_cost = tensor_redistribution.backward_comm_cost();
double computation_cost = tensor_redistribution.computation_cost();
double mem_cost = tensor_redistribution.memory_cost();
// Now AllGather, ReduceScatter, AlltoAll don't support bool type
MS_EXCEPTION_IF_NULL(type);
if ((type->type_id() == kNumberTypeBool) && (comm_cost > 0)) {
computation_cost = INF;
comm_cost = INF;
MS_LOG(WARNING) << "Communication Operators don't support bool dtype!";
}
*cost = std::make_shared<Cost>(type_length * computation_cost, type_length * comm_cost);
(*cost)->communication_without_parameter_ = type_length * comm_cost;
(*cost)->communication_with_partial_para_ =
(*cost)->communication_without_parameter_ +
COST_MODEL_GAMMA * ((*cost)->communication_cost_ - (*cost)->communication_without_parameter_);
(*cost)->communication_redis_forward_ = type_length * forward_comm_cost;
(*cost)->communication_redis_backward_ = type_length * backward_comm_cost;
(*cost)->memory_with_reuse_ = mem_cost;
return Status::SUCCESS;
}
CostPtrList Edge::GetCostList(StrategyPtr output_str, StrategyPtr input_str) {
CostPtrKey ck = {output_str, input_str};
CostPtrList result;
if (cost_map_.find(ck) != cost_map_.end()) {
return cost_map_.at(ck);
}
return result;
}
CostPtrList Edge::CreateEdgeEliminationCostList(const StrategyPtr &output_st_ptr, const std::vector<EdgePtr> &edges,
const StrategyPtr &input_st_ptr) {
std::function<CostPtrList(EdgePtr)> LocalGetCostList = [&](const EdgePtr &edge) {
MS_EXCEPTION_IF_NULL(edge);
return edge->GetCostList(output_st_ptr, input_st_ptr);
};
CostPtrList result;
std::vector<CostPtrList> all_cost_list;
all_cost_list.resize(edges.size());
(void)std::transform(edges.begin(), edges.end(), all_cost_list.begin(), LocalGetCostList);
CostPtrList selected_cost_list(all_cost_list.size(), nullptr);
std::function<void(size_t, double, double, double, double)> recursive =
[&](size_t k, double computation, double memory, double communication, double communication_without_para) {
if (k == edges.size()) {
auto decision = std::make_shared<EdgeEliminationDecision>(selected_cost_list);
CostPtr new_cost = std::make_shared<Cost>(computation, communication);
MS_EXCEPTION_IF_NULL(new_cost);
new_cost->communication_without_parameter_ = communication_without_para;
new_cost->communication_with_partial_para_ =
communication_without_para + COST_MODEL_GAMMA * (communication - communication_without_para);
new_cost->memory_with_reuse_ = memory;
new_cost->decision_ptr_ = decision;
result.push_back(new_cost);
return;
}
for (auto &c : all_cost_list[k]) {
MS_EXCEPTION_IF_NULL(c);
selected_cost_list[k] = c;
recursive(k + 1, computation + c->computation_cost_, memory + c->memory_with_reuse_,
communication + c->communication_cost_,
communication_without_para + c->communication_without_parameter_);
}
};
recursive(0, 0.0, 0.0, 0.0, 0.0);
SimplifyForDreasingCommunicationWithPartialPara(&result);
return result;
}
void Edge::EdgeEliminationSetNewCost(OperatorInfoPtr, const std::vector<EdgePtr> &edges, OperatorInfoPtr) {
bool valid = false;
for (const auto &output_pair : pre_op_output_) {
StrategyPtr output_st_ptr = output_pair.first;
for (const auto &input_pair : next_op_input_) {
StrategyPtr input_st_ptr = input_pair.first;
CostPtrList clist = CreateEdgeEliminationCostList(output_st_ptr, edges, input_st_ptr);
CostPtrKey key = {output_st_ptr, input_st_ptr};
cost_map_[key] = clist;
if ((!valid) && (!clist.empty())) {
valid = true;
}
}
}
if (!valid) {
MS_LOG(EXCEPTION) << "Creating edge: " << edge_name_ << " failed.";
}
}
void Edge::CreateOpEliminationSubCostList(StrategyPtr op_strategy, const CostPtrList &left_cost_list,
const CostPtrList &middle_cost_list, const CostPtrList &right_cost_list,
CostPtrList *ret_cost_list) {
for (auto &left_cost : left_cost_list) {
MS_EXCEPTION_IF_NULL(left_cost);
for (auto &middle_cost : middle_cost_list) {
MS_EXCEPTION_IF_NULL(middle_cost);
for (auto &right_cost : right_cost_list) {
MS_EXCEPTION_IF_NULL(right_cost);
double computation =
left_cost->computation_cost_ + middle_cost->computation_cost_ + right_cost->computation_cost_;
double communication =
left_cost->communication_cost_ + middle_cost->communication_cost_ + right_cost->communication_cost_;
double communication_without_para = left_cost->communication_without_parameter_ +
middle_cost->communication_without_parameter_ +
right_cost->communication_without_parameter_;
double memory_cost =
left_cost->memory_with_reuse_ + middle_cost->memory_with_reuse_ + right_cost->memory_with_reuse_;
auto decision = std::make_shared<OpEliminationDecision>(op_strategy, left_cost, middle_cost, right_cost);
auto cost = std::make_shared<Cost>(computation, communication, decision);
MS_EXCEPTION_IF_NULL(cost);
cost->communication_without_parameter_ = communication_without_para;
cost->communication_with_partial_para_ =
communication_without_para + COST_MODEL_GAMMA * (communication - communication_without_para);
cost->memory_with_reuse_ = memory_cost;
ret_cost_list->emplace_back(std::move(cost));
}
}
}
}
CostPtrList Edge::CreateOpEliminationCostList(const EdgePtr &e1, const StrategyPtr &output_st_ptr,
const OperatorInfoPtr &op, const EdgePtr &e2,
const StrategyPtr &input_st_ptr) {
MS_EXCEPTION_IF_NULL(op);
MS_EXCEPTION_IF_NULL(e1);
MS_EXCEPTION_IF_NULL(e2);
CostPtrList result;
for (const auto &op_strategy : op->GetStrategyCost()) {
MS_EXCEPTION_IF_NULL(op_strategy);
auto middle_strategy = op_strategy->strategy_ptr;
CreateOpEliminationSubCostList(middle_strategy, e1->GetCostList(output_st_ptr, middle_strategy),
op_strategy->cost_list, e2->GetCostList(middle_strategy, input_st_ptr), &result);
}
SimplifyForDreasingCommunicationWithPartialPara(&result);
return result;
}
void Edge::OpEliminationSetNewCost(const EdgePtr &e1, const OperatorInfoPtr &op, const EdgePtr &e2) {
bool valid = false;
for (const auto &output_pair : pre_op_output_) {
StrategyPtr output_st_ptr = output_pair.first;
for (const auto &input_pair : next_op_input_) {
StrategyPtr input_st_ptr = input_pair.first;
CostPtrList clist = CreateOpEliminationCostList(e1, output_st_ptr, op, e2, input_st_ptr);
CostPtrKey key = {output_st_ptr, input_st_ptr};
cost_map_[key] = clist;
if ((!valid) && (!clist.empty())) {
valid = true;
}
}
}
if (!valid) {
MS_LOG(EXCEPTION) << "Creating edge: " << edge_name_ << " failed.";
}
}
Status Edge::CalculateMemoryCost() {
if (is_output_parameter_involve_ == -1) {
MS_LOG(ERROR) << "is_output_parameter_involve_ is unset.";
return FAILED;
}
if (is_output_parameter_involve_ == 0) {
// In this case, it is sure that the tensor redistribution along this edge is NOT parameter-involved, thus it is
// unnecessary to keep them in memory.
for (auto &cost_kv : cost_map_) {
auto &cost_v = cost_kv.second;
if (!cost_v.empty()) {
cost_v[0]->memory_with_reuse_ = 0;
}
}
}
return SUCCESS;
}
} // namespace parallel
} // namespace mindspore