p04896132
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

merge master code to r0.5

This commit is contained in:
hongxing 2020-06-26 13:59:16 +02:00
parent fe1d4ca3bd
commit 300dd2971c
9 changed files with 340 additions and 225 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

125
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_cost.cc
View File

@ -28,7 +28,6 @@
namespace mindspore { namespace mindspore {
namespace parallel { namespace parallel {
#define DOUBLE_MAX (std::numeric_limits<double>::max)()
// Compute redistributed cost // Compute redistributed cost
double CostRedis(const Graph::NodeType &node, double CostRedis(const Graph::NodeType &node,
@ -621,75 +620,50 @@ StrategyRec CostCommon::ChoseStr(const std::vector<double> &cost_op, StrategyRec
break; break;
default: default:
MS_LOG(EXCEPTION) << "Failure: CostBiasAdd failed."; MS_LOG(EXCEPTION) << "Failure: Common failed.";
} }
return str; return str;
} }
// Get weight for BN // Get optimal strategy for BatchParallel OPs
double CostBatchNorm::GetMinCostIn(const OperatorRec &op) { StrategyRec CostBatchParallel::GetOptimalStr(const Graph::NodeType &node) {
int tensor = static_cast<int>(op.arguments[0].tensor_shape.shape_h * op.arguments[0].tensor_str.str_h) *
static_cast<int>(op.arguments[0].tensor_shape.shape_n * op.arguments[0].tensor_str.str_n) *
static_cast<int>(op.arguments[0].tensor_shape.shape_w * op.arguments[0].tensor_str.str_w) *
static_cast<int>(op.arguments[0].tensor_shape.shape_c * op.arguments[0].tensor_str.str_c);
std::vector<double> cost_in;
cost_in.push_back(StrDimB(tensor) * 1.2);
cost_in.push_back(DOUBLE_MAX);
cost_in.push_back(StrDimH(tensor) * 1.2);
cost_in.push_back(StrDimW(tensor) * 1.2);
return *min_element(cost_in.begin(), cost_in.end());
}
// Get optimal strategy for BN
StrategyRec CostBatchNorm::GetOptimalStr(const Graph::NodeType &node,
const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy,
const Graph &graph) {
const OperatorRec &op = node.apply; const OperatorRec &op = node.apply;
int tensor_n = static_cast<int>(op.arguments[0].tensor_shape.shape_n * op.arguments[0].tensor_str.str_n);
int tensor_filter_n = static_cast<int>(op.arguments[1].tensor_shape.shape_n * op.arguments[1].tensor_str.str_n); int tensor_c = static_cast<int>(op.arguments[0].tensor_shape.shape_c * op.arguments[0].tensor_str.str_c);
int tensor_filter_c = static_cast<int>(op.arguments[1].tensor_shape.shape_c * op.arguments[1].tensor_str.str_c); int tensor_h = static_cast<int>(op.arguments[0].tensor_shape.shape_h * op.arguments[0].tensor_str.str_h);
int tensor_filter_h = static_cast<int>(op.arguments[1].tensor_shape.shape_h * op.arguments[1].tensor_str.str_h); int tensor_w = static_cast<int>(op.arguments[0].tensor_shape.shape_w * op.arguments[0].tensor_str.str_w);
int tensor_filter_w = static_cast<int>(op.arguments[1].tensor_shape.shape_w * op.arguments[1].tensor_str.str_w);
int tensor_filter = tensor_filter_h * tensor_filter_w * tensor_filter_n * tensor_filter_c;
int output_tensor_h = static_cast<int>(node.tensor_parm.tensor_shape.shape_h * node.tensor_parm.tensor_str.str_h);
int output_tensor_w = static_cast<int>(node.tensor_parm.tensor_shape.shape_w * node.tensor_parm.tensor_str.str_w);
int output_tensor_n = static_cast<int>(node.tensor_parm.tensor_shape.shape_n * node.tensor_parm.tensor_str.str_n);
std::vector<double> cost_op; std::vector<double> cost_op;
std::vector<std::vector<float>> mode;
if (output_tensor_n < 2 || output_tensor_n % 2 != 0) { if (tensor_n < 2 || tensor_n % 2 != 0) {
cost_op.push_back(DOUBLE_MAX); cost_op.push_back(DOUBLE_MAX);
} else { } else {
cost_op.push_back(StrDimB(tensor_filter) + CostRedis(node, node_name_to_strategy, cost_op.push_back(cost_in_);
mode = {{0.5, 1, 1, 1}, {1, 1, 1, 1}, {0.5, 1, 1, 1}}, graph));
} }
cost_op.push_back(DOUBLE_MAX); if (tensor_c < 2 || tensor_c % 2 != 0) {
if (output_tensor_h < 2 || output_tensor_h % 2 != 0) {
cost_op.push_back(DOUBLE_MAX); cost_op.push_back(DOUBLE_MAX);
} else { } else {
cost_op.push_back(StrDimH(tensor_filter) + CostRedis(node, node_name_to_strategy, cost_op.push_back(cost_in_);
mode = {{1, 1, 0.5, 1}, {1, 1, 1, 1}, {1, 1, 0.5, 1}}, graph));
} }
if (output_tensor_w < 2 || output_tensor_w % 2 != 0) { if (tensor_h < 2 || tensor_h % 2 != 0) {
cost_op.push_back(DOUBLE_MAX); cost_op.push_back(DOUBLE_MAX);
} else { } else {
cost_op.push_back(StrDimW(tensor_filter) + CostRedis(node, node_name_to_strategy, cost_op.push_back(cost_in_);
mode = {{1, 1, 1, 0.5}, {1, 1, 1, 1}, {1, 1, 1, 0.5}}, graph)); }
if (tensor_w < 2 || tensor_w % 2 != 0) {
cost_op.push_back(DOUBLE_MAX);
} else {
cost_op.push_back(cost_in_);
} }
return ChoseStr(cost_op, node.apply.str); return ChoseStr(cost_op, node.apply.str);
} }
// Chose strategy for BatchNorm // Chose strategy for BatchParallel op
StrategyRec CostBatchNorm::ChoseStr(const std::vector<double> &cost_op, StrategyRec str) { StrategyRec CostBatchParallel::ChoseStr(const std::vector<double> &cost_op, StrategyRec str) {
uint64_t min_position = min_element(cost_op.begin(), cost_op.end()) - cost_op.begin(); uint64_t min_position = min_element(cost_op.begin(), cost_op.end()) - cost_op.begin();
if (cost_op[min_position] > (DOUBLE_MAX - 0.1)) { if (cost_op[min_position] > (DOUBLE_MAX - 0.1)) {
return str; return str;
@ -700,36 +674,75 @@ StrategyRec CostBatchNorm::ChoseStr(const std::vector<double> &cost_op, Strategy
str.inputTensor[0].str_n /= 2.0; str.inputTensor[0].str_n /= 2.0;
str.outputTensor.str_n /= 2.0; str.outputTensor.str_n /= 2.0;
str.cut_counter += 1; str.cut_counter += 1;
str.cost = str.cost + cost_in_b_; str.cost = str.cost + cost_in_;
break; break;
case 1: case 1:
str.inputTensor[0].str_c /= 2.0; str.inputTensor[0].str_c /= 2.0;
str.inputTensor[1].str_c /= 2.0;
str.inputTensor[2].str_c /= 2.0;
str.inputTensor[3].str_c /= 2.0;
str.inputTensor[4].str_c /= 2.0;
str.outputTensor.str_c /= 2.0; str.outputTensor.str_c /= 2.0;
str.cut_counter += 1; str.cut_counter += 1;
str.cost = str.cost + cost_in_c_; str.cost = str.cost + cost_in_;
break; break;
case 2: case 2:
str.inputTensor[0].str_h /= 2.0; str.inputTensor[0].str_h /= 2.0;
str.outputTensor.str_h /= 2.0; str.outputTensor.str_h /= 2.0;
str.cut_counter += 1; str.cut_counter += 1;
str.cost = str.cost + cost_in_h_; str.cost = str.cost + cost_in_;
break; break;
case 3: case 3:
str.inputTensor[0].str_w /= 2.0; str.inputTensor[0].str_w /= 2.0;
str.outputTensor.str_w /= 2.0; str.outputTensor.str_w /= 2.0;
str.cut_counter += 1; str.cut_counter += 1;
str.cost = str.cost + cost_in_w_; str.cost = str.cost + cost_in_;
break; break;
default: default:
MS_LOG(EXCEPTION) << "Failure: CostBatchNorm failed."; MS_LOG(EXCEPTION) << "Failure: CostBatchParallel failed.";
}
return str;
}
// Chose strategy for CostSoftmaxCrossEntropyWithLogits
StrategyRec CostSoftmaxCrossEntropyWithLogits::ChoseStr(const std::vector<double> &cost_op, StrategyRec str) {
uint64_t min_position = min_element(cost_op.begin(), cost_op.end()) - cost_op.begin();
if (cost_op[min_position] > (DOUBLE_MAX - 0.1)) {
return str;
}
switch (min_position) {
case 0:
str.inputTensor[0].str_n /= 2.0;
str.inputTensor[1].str_n /= 2.0;
str.cut_counter += 1;
str.cost = str.cost + cost_in_;
break;
case 1:
str.inputTensor[0].str_c /= 2.0;
str.inputTensor[1].str_c /= 2.0;
str.cut_counter += 1;
str.cost = str.cost + cost_in_;
break;
case 2:
str.inputTensor[0].str_h /= 2.0;
str.inputTensor[1].str_h /= 2.0;
str.outputTensor.str_w /= 2.0;
str.cut_counter += 1;
str.cost = str.cost + cost_in_;
break;
case 3:
str.inputTensor[0].str_w /= 2.0;
str.inputTensor[1].str_w /= 2.0;
str.cut_counter += 1;
str.cost = str.cost + cost_in_;
break;
default:
MS_LOG(EXCEPTION) << "Failure: CostSoftmax failed.";
} }
return str; return str;
} }

54
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_cost.h
View File

@ -28,6 +28,8 @@
namespace mindspore { namespace mindspore {
namespace parallel { namespace parallel {
#define DOUBLE_MAX (std::numeric_limits<double>::max)()
double CostRedis(const Graph::NodeType &node, double CostRedis(const Graph::NodeType &node,
const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy, const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy,
const std::vector<std::vector<float>> &mode, const Graph &graph); const std::vector<std::vector<float>> &mode, const Graph &graph);
@ -195,7 +197,6 @@ class CostTensorAdd : public CostCommon {
}; };
// all the following operation are element-wise and have the same cost // all the following operation are element-wise and have the same cost
class CostOneHot : public CostCommon {};
class CostReLU : public CostCommon {}; class CostReLU : public CostCommon {};
class CostLog : public CostCommon {}; class CostLog : public CostCommon {};
class CostExp : public CostCommon {}; class CostExp : public CostCommon {};
@ -206,50 +207,27 @@ class CostDiv : public CostCommon {};
class CostSqueeze : public CostCommon {}; class CostSqueeze : public CostCommon {};
class CostCast : public CostCommon {}; class CostCast : public CostCommon {};
// class BatchNorm is used to compute the cost of BatchNorm operator. // class BatchParallel is used to compute the cost of BatchParallel operator.
class CostBatchNorm { class CostBatchParallel {
public: public:
StrategyRec GetOptimalStr(const Graph::NodeType &node, virtual StrategyRec GetOptimalStr(const Graph::NodeType &node);
const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy,
const Graph &graph);
double GetMinCostIn(const OperatorRec &op); virtual double GetMaxCostIn() const { return DOUBLE_MAX; }
private: protected:
double StrDimB(int32_t Tensor) { virtual StrategyRec ChoseStr(const std::vector<double> &cost_op, StrategyRec str);
cost_in_b_ = (static_cast<double>(Tensor) * 4.0) / 2.0;
return cost_in_b_; double cost_in_ = 0;
} }; // class BatchParallel is used to compute the cost of BatchParallel operator.
double StrDimC() { class CostBatchNorm : public CostBatchParallel {};
cost_in_c_ = 0.0; class CostOneHot : public CostBatchParallel {};
class CostPRelu : public CostBatchParallel {};
return cost_in_c_; class CostSoftmax : public CostBatchParallel {};
}
double StrDimH(int32_t Tensor) {
cost_in_h_ = (static_cast<double>(Tensor) * 4.0) / 2.0;
return cost_in_h_;
}
double StrDimW(int32_t Tensor) {
cost_in_w_ = (static_cast<double>(Tensor) * 4.0) / 2.0;
return cost_in_w_;
}
class CostSoftmaxCrossEntropyWithLogits : public CostBatchParallel {
StrategyRec ChoseStr(const std::vector<double> &cost_op, StrategyRec str); StrategyRec ChoseStr(const std::vector<double> &cost_op, StrategyRec str);
};
double cost_in_b_ = 0;
double cost_in_c_ = 0;
double cost_in_h_ = 0;
double cost_in_w_ = 0;
}; // class BatchNorm is used to compute the cost of BatchNorm operator.
} // namespace parallel } // namespace parallel
} // namespace mindspore } // namespace mindspore
#endif // PARALLEL_AUTO_PARALLEL_REC_COST_H_ #endif // PARALLEL_AUTO_PARALLEL_REC_COST_H_

214
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_generate_strategy.cc
View File

@ -28,10 +28,10 @@
namespace mindspore { namespace mindspore {
namespace parallel { namespace parallel {
void GenerateStrategy(std::shared_ptr<Graph> graph, const std::vector<std::shared_ptr<OperatorInfo>> &ops, void GenerateStrategy(const std::shared_ptr<Graph> &graph, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::shared_ptr<std::vector<std::vector<size_t>>> eli_list, const std::shared_ptr<std::vector<std::vector<size_t>>> &eli_list,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> index_list) { const std::shared_ptr<std::vector<size_t>> &index_list) {
MS_EXCEPTION_IF_NULL(graph); MS_EXCEPTION_IF_NULL(graph);
MS_EXCEPTION_IF_NULL(eli_list); MS_EXCEPTION_IF_NULL(eli_list);
MS_EXCEPTION_IF_NULL(index_list); MS_EXCEPTION_IF_NULL(index_list);
@ -127,25 +127,6 @@ std::vector<std::vector<int32_t>> PrepareMatMul(const std::shared_ptr<Graph> &gr
return strategies; return strategies;
} }
std::vector<std::vector<int32_t>> PreparePReLU(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops) {
std::vector<std::vector<int32_t>> strategies = MakeDataParallelStrategy(graph, ops, iter_graph, iter_ops);
strategies[1][0] = 1;
return strategies;
}
std::vector<std::vector<int32_t>> PrepareBatchNorm(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops) {
std::vector<std::vector<int32_t>> strategies = MakeDataParallelStrategy(graph, ops, iter_graph, iter_ops);
for (size_t i = 1; i < strategies.size(); i++) {
strategies[i][0] = strategies[0][1];
}
strategies[1][0] = 1;
return strategies;
}
std::vector<std::vector<int32_t>> PrepareBiasAdd(const std::shared_ptr<std::vector<int32_t>> &s) { std::vector<std::vector<int32_t>> PrepareBiasAdd(const std::shared_ptr<std::vector<int32_t>> &s) {
std::vector<std::vector<int32_t>> strategies; std::vector<std::vector<int32_t>> strategies;
strategies.push_back(*s); strategies.push_back(*s);
@ -155,10 +136,29 @@ std::vector<std::vector<int32_t>> PrepareBiasAdd(const std::shared_ptr<std::vect
return strategies; return strategies;
} }
std::vector<std::vector<int32_t>> PrepareOneHot(const std::shared_ptr<std::vector<int32_t>> &s) { std::vector<std::vector<int32_t>> PrepareOneHot(const std::shared_ptr<Graph> &graph,
std::vector<std::vector<int32_t>> strategies; const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops) {
std::vector<std::vector<int32_t>> strategies = MakeRecSearchStrategy(graph, ops, iter_graph, iter_ops);
int32_t axis = -1;
auto iter = ops[iter_ops]->attrs().find(AXIS);
if (iter != ops[iter_ops]->attrs().end()) {
MS_EXCEPTION_IF_NULL(iter->second);
if (iter->second->isa<Int32Imm>()) {
axis = iter->second->cast<Int32ImmPtr>()->value();
} else {
MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": The value of axis is not int.";
}
}
if (axis == -1) {
strategies[0][0] = strategies[0][1];
strategies[0][1] = 1;
graph->nodes[iter_graph].tensor_parm.tensor_str.str_h = graph->nodes[iter_graph].tensor_parm.tensor_str.str_w;
graph->nodes[iter_graph].tensor_parm.tensor_str.str_w = 1.0;
}
std::vector<int32_t> s_empty = {}; std::vector<int32_t> s_empty = {};
strategies.push_back(*s);
strategies.push_back(s_empty); strategies.push_back(s_empty);
strategies.push_back(s_empty); strategies.push_back(s_empty);
return strategies; return strategies;
@ -170,6 +170,32 @@ std::vector<std::vector<int32_t>> PrepareGatherV2(const std::shared_ptr<std::vec
return strategies; return strategies;
} }
std::vector<std::vector<int32_t>> PrepareL2Normalize(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, std::vector<int32_t> s) {
int32_t axis = 0;
auto iter = ops[iter_ops]->attrs().find(AXIS);
if (iter != ops[iter_ops]->attrs().end()) {
MS_EXCEPTION_IF_NULL(iter->second);
if (iter->second->isa<Int32Imm>()) {
axis = iter->second->cast<Int32ImmPtr>()->value();
} else {
MS_LOG(EXCEPTION) << ops[iter_ops]->name() << " : The value of axis is not int.";
}
}
int32_t axis_index = axis;
if (axis < 0) {
size_t input_dim = ops[iter_ops]->inputs_tensor_info()[0].shape().size();
axis_index = static_cast<int32_t>(input_dim) + axis;
}
s[IntToSize(axis_index)] = 1;
std::vector<std::vector<int32_t>> strategies;
strategies.push_back(s);
return strategies;
}
std::vector<std::vector<int32_t>> MakeRecSearchStrategy(const std::shared_ptr<Graph> &graph, std::vector<std::vector<int32_t>> MakeRecSearchStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops) { const size_t iter_graph, const size_t iter_ops) {
@ -209,7 +235,7 @@ std::vector<std::vector<int32_t>> MakeRecSearchStrategy(const std::shared_ptr<Gr
} else if (output_size == 0) { } else if (output_size == 0) {
s = {}; s = {};
} else { } else {
MS_LOG(ERROR) << "Tensor's output size is unexcepted."; MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": Tensor's output size is unexcepted.";
} }
strategies.push_back(s); strategies.push_back(s);
} }
@ -229,7 +255,7 @@ std::vector<std::vector<int32_t>> MakeDataParallelStrategy(const std::shared_ptr
StrategyPtr origin_strategy = ops[iter_ops]->strategy(); StrategyPtr origin_strategy = ops[iter_ops]->strategy();
std::vector<std::vector<int32_t>> strategies; std::vector<std::vector<int32_t>> strategies;
size_t max_device_num = g_device_manager->DeviceNum(); size_t max_device_num = g_device_manager->DeviceNum();
size_t target_tensor_batch = ops[iter_ops]->outputs_tensor_info()[0].shape()[0]; size_t target_tensor_batch = ops[iter_ops]->inputs_tensor_info()[0].shape()[0];
for (size_t iter_op_inputs = 0; iter_op_inputs < ops[iter_ops]->inputs_tensor_info().size(); iter_op_inputs++) { for (size_t iter_op_inputs = 0; iter_op_inputs < ops[iter_ops]->inputs_tensor_info().size(); iter_op_inputs++) {
if (iter_op_inputs >= origin_strategy->GetInputDim().size()) { if (iter_op_inputs >= origin_strategy->GetInputDim().size()) {
MS_LOG(EXCEPTION) << "Failure: Strategy's InputDim out of range."; MS_LOG(EXCEPTION) << "Failure: Strategy's InputDim out of range.";
@ -244,8 +270,10 @@ std::vector<std::vector<int32_t>> MakeDataParallelStrategy(const std::shared_ptr
} else { } else {
s.push_back(1); s.push_back(1);
} }
} else if (input_size == 0) {
s = {};
} else { } else {
MS_LOG(ERROR) << "Tensor's shape is unknown."; MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": Tensor's shape is unknown.";
} }
} }
strategies.push_back(s); strategies.push_back(s);
@ -285,25 +313,20 @@ std::vector<std::vector<int32_t>> PrepareStrategy(const std::shared_ptr<Graph> &
if (type == MATMUL) { if (type == MATMUL) {
return PrepareMatMul(graph, ops, iter_graph, iter_ops); return PrepareMatMul(graph, ops, iter_graph, iter_ops);
} else if (type == PRELU) { } else if (type == ONEHOT) {
return PreparePReLU(graph, ops, iter_graph, iter_ops); return PrepareOneHot(graph, ops, iter_graph, iter_ops);
} else if (type == BATCH_NORM) {
return PrepareBatchNorm(graph, ops, iter_graph, iter_ops);
} else if (type == SOFTMAX || type == LOG_SOFTMAX || type == SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS ||
type == SOFTMAX_CROSS_ENTROPY_WITH_LOGITS) {
return MakeDataParallelStrategy(graph, ops, iter_graph, iter_ops);
} else { } else {
return MakeRecSearchStrategy(graph, ops, iter_graph, iter_ops); return MakeRecSearchStrategy(graph, ops, iter_graph, iter_ops);
} }
} }
void GeneratePartitionedOperatorStrategy(const std::shared_ptr<Graph> graph, void GeneratePartitionedOperatorStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::shared_ptr<std::vector<size_t>> index_list) { const std::shared_ptr<std::vector<size_t>> &index_list) {
for (size_t iter_ops = 0; iter_ops < (size_t)index_list->size(); iter_ops++) { for (size_t iter_ops = 0; iter_ops < (size_t)index_list->size(); iter_ops++) {
std::vector<std::vector<int32_t>> strategies; std::vector<std::vector<int32_t>> strategies;
size_t iter_graph = index_list->at(iter_ops); size_t iter_graph = index_list->at(iter_ops);
if (iter_graph != SIZE_MAX) { if (iter_graph != SIZE_MAX && ops[iter_ops]->type() != GET_NEXT) {
strategies = PrepareStrategy(graph, ops, iter_graph, iter_ops); strategies = PrepareStrategy(graph, ops, iter_graph, iter_ops);
} }
StrategyPtr sp = std::make_shared<Strategy>(0, strategies); StrategyPtr sp = std::make_shared<Strategy>(0, strategies);
@ -328,7 +351,7 @@ size_t FindIndexOfOperatorIncoming(const std::vector<std::vector<std::string>> &
return incoming_op_index; return incoming_op_index;
} }
std::vector<int32_t> CopyIncomingOperatorOutputStrategy(const std::shared_ptr<Graph> graph, std::vector<int32_t> CopyIncomingOperatorOutputStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, const size_t iter_graph) { const size_t iter_ops, const size_t iter_graph) {
std::vector<int32_t> s; std::vector<int32_t> s;
@ -348,7 +371,7 @@ std::vector<int32_t> CopyIncomingOperatorOutputStrategy(const std::shared_ptr<Gr
s.push_back(1 / graph->nodes[iter_graph].tensor_parm.tensor_str.str_h); s.push_back(1 / graph->nodes[iter_graph].tensor_parm.tensor_str.str_h);
s.push_back(1 / graph->nodes[iter_graph].tensor_parm.tensor_str.str_w); s.push_back(1 / graph->nodes[iter_graph].tensor_parm.tensor_str.str_w);
} else { } else {
MS_LOG(ERROR) << "Tensor's shape is unknown."; MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": Tensor's shape is unknown.";
} }
break; break;
} }
@ -358,7 +381,8 @@ std::vector<int32_t> CopyIncomingOperatorOutputStrategy(const std::shared_ptr<Gr
std::vector<int32_t> PrepareIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> PrepareIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t incoming_op_index) { const size_t incoming_op_index) {
std::vector<int32_t> s; std::vector<int32_t> s;
if (ops[incoming_op_index]->type() == RESHAPE || ops[incoming_op_index]->type() == GATHERV2) { if (ops[incoming_op_index]->type() == RESHAPE || ops[incoming_op_index]->type() == GATHERV2 ||
ops[incoming_op_index]->type() == TRANSPOSE) {
return s; return s;
} }
auto strategy = ops[incoming_op_index]->selected_strategy(); auto strategy = ops[incoming_op_index]->selected_strategy();
@ -426,13 +450,23 @@ std::vector<int32_t> ModifyStrategyIfSqueezeIncoming(const std::vector<std::shar
return s_Squeeze; return s_Squeeze;
} }
bool GetKeepDims(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops) {
bool keepdims = false;
auto keep_dims_iter = ops[iter_ops]->attrs().find(KEEP_DIMS);
if (keep_dims_iter == ops[iter_ops]->attrs().end()) {
MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": Don't have attr keep_dims.";
}
MS_EXCEPTION_IF_NULL(keep_dims_iter->second);
if (!keep_dims_iter->second->isa<BoolImm>()) {
MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": Keep_dims is not a bool.";
}
keepdims = keep_dims_iter->second->cast<BoolImmPtr>()->value();
return keepdims;
}
std::vector<int32_t> GetDimList(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops) { std::vector<int32_t> GetDimList(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops) {
std::vector<int32_t> dim_list; std::vector<int32_t> dim_list;
bool keep_dims; bool keep_dims = GetKeepDims(ops, iter_ops);
if (!ops[iter_ops]->attrs().find(KEEP_DIMS)->second->isa<BoolImm>()) {
MS_LOG(EXCEPTION) << "Failure: Parameter keep_dims is not a boolean value." << std::endl;
}
keep_dims = ops[iter_ops]->attrs().find(KEEP_DIMS)->second->cast<BoolImmPtr>()->value();
if (keep_dims != false) { if (keep_dims != false) {
return dim_list; return dim_list;
} }
@ -478,6 +512,62 @@ std::vector<int32_t> ModifyStrategyIfReduceIncoming(const std::vector<std::share
return s_Reduce; return s_Reduce;
} }
std::vector<int32_t> GetDimListFromAttrs(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops) {
std::vector<int32_t> dim_list;
auto iter = ops[iter_ops]->attrs().find(AXIS);
if (iter == ops[iter_ops]->attrs().end()) {
MS_LOG(EXCEPTION) << ops[iter_ops]->name() << ": Don't have attr axis.";
}
auto input_dim = ops[iter_ops]->inputs_tensor_info()[0].shape().size();
MS_EXCEPTION_IF_NULL(iter->second);
if (iter->second->isa<ValueTuple>()) {
auto attr_axis = GetValue<std::vector<int>>(iter->second);
if (attr_axis.empty()) {
for (size_t i = 0; i < input_dim; ++i) {
dim_list.push_back(SizeToInt(i));
}
} else {
for (auto &axis : attr_axis) {
axis < 0 ? dim_list.push_back(axis + SizeToInt(input_dim)) : dim_list.push_back(axis);
}
}
} else if (iter->second->isa<Int32Imm>()) {
int axis = GetValue<int>(iter->second);
axis < 0 ? dim_list.push_back(axis + SizeToInt(input_dim)) : dim_list.push_back(axis);
} else {
MS_LOG(EXCEPTION) << "Axis type is invalid.";
}
return dim_list;
}
std::vector<int32_t> ModifyStrategyIfArgIncoming(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t incoming_op_index, std::vector<int32_t> s) {
bool keepdims = GetKeepDims(ops, incoming_op_index);
if (keepdims) {
return s;
}
std::vector<int32_t> s_Arg;
std::vector<int32_t> axis_list;
for (size_t i = 0; i < s.size(); i++) {
axis_list.push_back(i);
}
auto dim_list = GetDimListFromAttrs(ops, incoming_op_index);
for (auto axis : dim_list) {
auto it = find(axis_list.begin(), axis_list.end(), axis);
if (it == axis_list.end()) {
MS_LOG(EXCEPTION) << "Failure: Can not find dimension indexes in Axis." << std::endl;
}
axis_list.erase(it);
}
for (size_t i = 0; i < (size_t)axis_list.size(); i++) {
s_Arg.push_back(s[axis_list[i]]);
}
return s_Arg;
}
std::vector<int32_t> CopyIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> CopyIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, const size_t incoming_op_index) { const size_t iter_ops, const size_t incoming_op_index) {
std::vector<int32_t> s; std::vector<int32_t> s;
@ -490,6 +580,9 @@ std::vector<int32_t> CopyIncomingOperatorInputStrategy(const std::vector<std::sh
ops[incoming_op_index]->type() == REDUCE_MIN || ops[incoming_op_index]->type() == REDUCE_MEAN) { ops[incoming_op_index]->type() == REDUCE_MIN || ops[incoming_op_index]->type() == REDUCE_MEAN) {
s = ModifyStrategyIfReduceIncoming(ops, incoming_op_index, s); s = ModifyStrategyIfReduceIncoming(ops, incoming_op_index, s);
} }
if (ops[incoming_op_index]->type() == ARGMAXWITHVALUE || ops[incoming_op_index]->type() == ARGMINWITHVALUE) {
s = ModifyStrategyIfArgIncoming(ops, incoming_op_index, s);
}
} }
return s; return s;
} }
@ -513,12 +606,12 @@ std::vector<std::vector<int32_t>> GenerateStrategiesFromStrategy(const std::vect
if (ops[iter_ops]->type() == BIAS_ADD) { if (ops[iter_ops]->type() == BIAS_ADD) {
return PrepareBiasAdd(s_ptr); return PrepareBiasAdd(s_ptr);
} }
if (ops[iter_ops]->type() == ONEHOT) {
return PrepareOneHot(s_ptr);
}
if (ops[iter_ops]->type() == GATHERV2) { if (ops[iter_ops]->type() == GATHERV2) {
return PrepareGatherV2(s_ptr); return PrepareGatherV2(s_ptr);
} }
if (ops[iter_ops]->type() == L2_NORMALIZE) {
return PrepareL2Normalize(ops, iter_ops, basic_stra);
}
for (size_t iter_op_inputs = 0; iter_op_inputs < (size_t)ops[iter_ops]->inputs_tensor_info().size(); for (size_t iter_op_inputs = 0; iter_op_inputs < (size_t)ops[iter_ops]->inputs_tensor_info().size();
iter_op_inputs++) { iter_op_inputs++) {
@ -544,11 +637,11 @@ std::vector<std::vector<int32_t>> GenerateStrategiesFromStrategy(const std::vect
return stra; return stra;
} }
void GenerateEliminatedOperatorStrategyForward(const std::shared_ptr<Graph> graph, void GenerateEliminatedOperatorStrategyForward(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> index_list, const std::shared_ptr<std::vector<size_t>> &index_list,
const std::shared_ptr<std::vector<size_t>> no_stra_op_list) { const std::shared_ptr<std::vector<size_t>> &no_stra_op_list) {
if (no_stra_op_list->size() == 0) { if (no_stra_op_list->size() == 0) {
return; return;
} }
@ -559,7 +652,7 @@ void GenerateEliminatedOperatorStrategyForward(const std::shared_ptr<Graph> grap
std::vector<std::vector<int32_t>> stra; std::vector<std::vector<int32_t>> stra;
std::vector<int32_t> s; std::vector<int32_t> s;
size_t incoming_op_index = FindIndexOfOperatorIncoming(input_tensor_names, iter_ops); size_t incoming_op_index = FindIndexOfOperatorIncoming(input_tensor_names, iter_ops);
if (incoming_op_index != SIZE_MAX && ops[iter_ops]->type() != ONEHOT) { if (incoming_op_index != SIZE_MAX) {
auto iter_graph = index_list->at(incoming_op_index); auto iter_graph = index_list->at(incoming_op_index);
if (iter_graph != SIZE_MAX) { if (iter_graph != SIZE_MAX) {
s = CopyIncomingOperatorOutputStrategy(graph, ops, iter_ops, iter_graph); s = CopyIncomingOperatorOutputStrategy(graph, ops, iter_ops, iter_graph);
@ -617,7 +710,8 @@ std::vector<int32_t> CopyOutgoingOperatorInputStrategy(const std::vector<std::sh
std::vector<int32_t> s; std::vector<int32_t> s;
if (ops[iter_ops]->type() == REDUCE_MAX || ops[iter_ops]->type() == REDUCE_MIN || if (ops[iter_ops]->type() == REDUCE_MAX || ops[iter_ops]->type() == REDUCE_MIN ||
ops[iter_ops]->type() == REDUCE_SUM || ops[iter_ops]->type() == REDUCE_MEAN || ops[iter_ops]->type() == RESHAPE || ops[iter_ops]->type() == REDUCE_SUM || ops[iter_ops]->type() == REDUCE_MEAN || ops[iter_ops]->type() == RESHAPE ||
ops[iter_ops]->type() == GATHERV2) { ops[iter_ops]->type() == GATHERV2 || ops[iter_ops]->type() == TRANSPOSE ||
ops[iter_ops]->type() == ARGMAXWITHVALUE || ops[iter_ops]->type() == ARGMINWITHVALUE) {
return s; return s;
} }
@ -640,7 +734,7 @@ std::vector<int32_t> CopyOutgoingOperatorInputStrategy(const std::vector<std::sh
} }
if (outgoing_op_index != SIZE_MAX && iter_op_inputs != SIZE_MAX) { if (outgoing_op_index != SIZE_MAX && iter_op_inputs != SIZE_MAX) {
for (size_t k = 0; k < ops[outgoing_op_index]->selected_strategy()->GetInputDim()[iter_op_inputs].size(); ++k) { for (size_t k = 0; k < ops[iter_ops]->outputs_tensor_info()[0].shape().size(); ++k) {
s.push_back(ops[outgoing_op_index]->selected_strategy()->GetInputDim()[iter_op_inputs][k]); s.push_back(ops[outgoing_op_index]->selected_strategy()->GetInputDim()[iter_op_inputs][k]);
} }
} }
@ -649,7 +743,7 @@ std::vector<int32_t> CopyOutgoingOperatorInputStrategy(const std::vector<std::sh
void GenerateEliminatedOperatorStrategyBackward(const std::vector<std::shared_ptr<OperatorInfo>> &ops, void GenerateEliminatedOperatorStrategyBackward(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> no_stra_op_list) { const std::shared_ptr<std::vector<size_t>> &no_stra_op_list) {
if (no_stra_op_list->size() == 0) { if (no_stra_op_list->size() == 0) {
return; return;
} }
@ -679,16 +773,16 @@ void GenerateEliminatedOperatorStrategyBackward(const std::vector<std::shared_pt
} }
} }
void GenerateRemainingOperatorStrategy(const std::shared_ptr<Graph> graph, void GenerateRemainingOperatorStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> index_list, const std::shared_ptr<std::vector<size_t>> &index_list,
const std::shared_ptr<std::vector<size_t>> no_stra_op_list) { const std::shared_ptr<std::vector<size_t>> &no_stra_op_list) {
if (no_stra_op_list->size() == 0) { if (no_stra_op_list->size() == 0) {
return; return;
} }
size_t no_stra_op_list_size; size_t no_stra_op_list_size = no_stra_op_list->size();
do { do {
no_stra_op_list_size = no_stra_op_list->size(); no_stra_op_list_size = no_stra_op_list->size();
GenerateEliminatedOperatorStrategyForward(graph, ops, input_tensor_names, index_list, no_stra_op_list); GenerateEliminatedOperatorStrategyForward(graph, ops, input_tensor_names, index_list, no_stra_op_list);

42
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_generate_strategy.h
View File

@ -27,22 +27,20 @@
namespace mindspore { namespace mindspore {
namespace parallel { namespace parallel {
void GenerateStrategy(std::shared_ptr<Graph> graph, const std::vector<std::shared_ptr<OperatorInfo>> &ops, void GenerateStrategy(const std::shared_ptr<Graph> &graph, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::shared_ptr<std::vector<std::vector<size_t>>> eli_list, const std::shared_ptr<std::vector<std::vector<size_t>>> &eli_list,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> index_list); const std::shared_ptr<std::vector<size_t>> &index_list);
std::vector<std::vector<int32_t>> PrepareMatMul(const std::shared_ptr<Graph> &graph, std::vector<std::vector<int32_t>> PrepareMatMul(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops); const size_t iter_graph, const size_t iter_ops);
std::vector<std::vector<int32_t>> PreparePReLU(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops);
std::vector<std::vector<int32_t>> PrepareBatchNorm(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops);
std::vector<std::vector<int32_t>> PrepareBiasAdd(const std::shared_ptr<std::vector<int32_t>> &s); std::vector<std::vector<int32_t>> PrepareBiasAdd(const std::shared_ptr<std::vector<int32_t>> &s);
std::vector<std::vector<int32_t>> PrepareOneHot(const std::shared_ptr<std::vector<int32_t>> &s); std::vector<std::vector<int32_t>> PrepareOneHot(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops);
std::vector<std::vector<int32_t>> PrepareGatherV2(const std::shared_ptr<std::vector<int32_t>> &s); std::vector<std::vector<int32_t>> PrepareGatherV2(const std::shared_ptr<std::vector<int32_t>> &s);
std::vector<std::vector<int32_t>> PrepareL2Normalize(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, std::vector<int32_t> s);
std::vector<std::vector<int32_t>> MakeRecSearchStrategy(const std::shared_ptr<Graph> &graph, std::vector<std::vector<int32_t>> MakeRecSearchStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops); const size_t iter_graph, const size_t iter_ops);
@ -52,12 +50,12 @@ std::vector<std::vector<int32_t>> MakeDataParallelStrategy(const std::shared_ptr
std::vector<std::vector<int32_t>> PrepareStrategy(const std::shared_ptr<Graph> &graph, std::vector<std::vector<int32_t>> PrepareStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_graph, const size_t iter_ops); const size_t iter_graph, const size_t iter_ops);
void GeneratePartitionedOperatorStrategy(const std::shared_ptr<Graph> graph, void GeneratePartitionedOperatorStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::shared_ptr<std::vector<size_t>> index_list); const std::shared_ptr<std::vector<size_t>> &index_list);
size_t FindIndexOfOperatorIncoming(const std::vector<std::vector<std::string>> &input_tensor_names, size_t FindIndexOfOperatorIncoming(const std::vector<std::vector<std::string>> &input_tensor_names,
const size_t iter_ops); const size_t iter_ops);
std::vector<int32_t> CopyIncomingOperatorOutputStrategy(const std::shared_ptr<Graph> graph, std::vector<int32_t> CopyIncomingOperatorOutputStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, const size_t iter_graph); const size_t iter_ops, const size_t iter_graph);
std::vector<int32_t> PrepareIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> PrepareIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
@ -65,19 +63,23 @@ std::vector<int32_t> PrepareIncomingOperatorInputStrategy(const std::vector<std:
std::vector<int32_t> GetAxisList(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const int iter_ops); std::vector<int32_t> GetAxisList(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const int iter_ops);
std::vector<int32_t> ModifyStrategyIfSqueezeIncoming(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> ModifyStrategyIfSqueezeIncoming(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t incoming_op_index, std::vector<int32_t> s); const size_t incoming_op_index, std::vector<int32_t> s);
bool GetKeepDims(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops);
std::vector<int32_t> GetDimList(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops); std::vector<int32_t> GetDimList(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops);
std::vector<int32_t> ModifyStrategyIfReduceIncoming(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> ModifyStrategyIfReduceIncoming(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t incoming_op_index, std::vector<int32_t> s); const size_t incoming_op_index, std::vector<int32_t> s);
std::vector<int32_t> GetDimListFromAttrs(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops);
std::vector<int32_t> ModifyStrategyIfArgIncoming(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t incoming_op_index, std::vector<int32_t> s);
std::vector<int32_t> CopyIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> CopyIncomingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, const size_t incoming_op_index); const size_t iter_ops, const size_t incoming_op_index);
std::vector<std::vector<int32_t>> GenerateStrategiesFromStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<std::vector<int32_t>> GenerateStrategiesFromStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, const size_t iter_ops,
std::vector<int32_t> basic_stra); std::vector<int32_t> basic_stra);
void GenerateEliminatedOperatorStrategyForward(std::shared_ptr<Graph> graph, void GenerateEliminatedOperatorStrategyForward(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> index_list, const std::shared_ptr<std::vector<size_t>> &index_list,
const std::shared_ptr<std::vector<size_t>> no_stra_op_list); const std::shared_ptr<std::vector<size_t>> &no_stra_op_list);
std::vector<int32_t> ModifyStrategyIfSqueezeOutgoing(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> ModifyStrategyIfSqueezeOutgoing(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const size_t iter_ops, std::vector<int32_t> s); const size_t iter_ops, std::vector<int32_t> s);
std::vector<int32_t> CopyOutgoingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::vector<int32_t> CopyOutgoingOperatorInputStrategy(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
@ -85,12 +87,12 @@ std::vector<int32_t> CopyOutgoingOperatorInputStrategy(const std::vector<std::sh
const size_t iter_ops); const size_t iter_ops);
void GenerateEliminatedOperatorStrategyBackward(const std::vector<std::shared_ptr<OperatorInfo>> &ops, void GenerateEliminatedOperatorStrategyBackward(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> no_stra_op_list); const std::shared_ptr<std::vector<size_t>> &no_stra_op_list);
void GenerateRemainingOperatorStrategy(const std::shared_ptr<Graph> graph, void GenerateRemainingOperatorStrategy(const std::shared_ptr<Graph> &graph,
const std::vector<std::shared_ptr<OperatorInfo>> &ops, const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names, const std::vector<std::vector<std::string>> &input_tensor_names,
const std::shared_ptr<std::vector<size_t>> index_list, const std::shared_ptr<std::vector<size_t>> &index_list,
const std::shared_ptr<std::vector<size_t>> no_stra_op_list); const std::shared_ptr<std::vector<size_t>> &no_stra_op_list);
} // namespace parallel } // namespace parallel
} // namespace mindspore } // namespace mindspore
#endif // PARALLEL_AUTO_PARALLEL_REC_GENERATE_STRATEGY_H_ #endif // PARALLEL_AUTO_PARALLEL_REC_GENERATE_STRATEGY_H_

4
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_graph.h
View File

@ -38,6 +38,7 @@ enum OperatorType {
kRecBiasAdd, kRecBiasAdd,
kRecSoftmax, kRecSoftmax,
kRecSparseSoftmaxCrossEntropyWithLogits, kRecSparseSoftmaxCrossEntropyWithLogits,
kRecSoftmaxCrossEntropyWithLogits,
kRecOneHot, kRecOneHot,
kRecLog, kRecLog,
kRecExp, kRecExp,
@ -49,7 +50,8 @@ enum OperatorType {
kRecCast, kRecCast,
kRecReduce, kRecReduce,
kRecPReLU, kRecPReLU,
kRecGatherV2 kRecGatherV2,
kRecArgWithValue
}; };
enum InfoType { kApplication, kConstant }; enum InfoType { kApplication, kConstant };

44
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_parse_graph.cc
View File

@ -40,7 +40,7 @@ const TensorParam MakeTensor(int n, int c, int h, int w) {
return tensor; return tensor;
} }
Graph::NodeType MakeNewOperator(std::vector<std::shared_ptr<OperatorInfo>> ops, size_t iter_ops) { Graph::NodeType MakeNewOperator(const std::vector<std::shared_ptr<OperatorInfo>> &ops, size_t iter_ops) {
Graph::NodeType NewOp; Graph::NodeType NewOp;
NewOp.name = ops[iter_ops]->name(); NewOp.name = ops[iter_ops]->name();
NewOp.info = InfoType::kApplication; NewOp.info = InfoType::kApplication;
@ -140,7 +140,7 @@ std::shared_ptr<Graph> ParseGraph(const std::vector<std::shared_ptr<OperatorInfo
return graph; return graph;
} }
void MakeEdge(const std::vector<std::vector<std::string>> &input_tensor_names, std::shared_ptr<Graph> graph) { void MakeEdge(const std::vector<std::vector<std::string>> &input_tensor_names, const std::shared_ptr<Graph> &graph) {
for (size_t iter_i = 0; iter_i < input_tensor_names.size(); iter_i++) { for (size_t iter_i = 0; iter_i < input_tensor_names.size(); iter_i++) {
for (size_t iter_j = 1; iter_j < input_tensor_names[iter_i].size(); iter_j++) { for (size_t iter_j = 1; iter_j < input_tensor_names[iter_i].size(); iter_j++) {
size_t head_node_index = GetIndexInInputTensorNames(input_tensor_names, input_tensor_names[iter_i][iter_j]); size_t head_node_index = GetIndexInInputTensorNames(input_tensor_names, input_tensor_names[iter_i][iter_j]);
@ -163,8 +163,8 @@ size_t GetIndexInInputTensorNames(const std::vector<std::vector<std::string>> &i
return SIZE_MAX; return SIZE_MAX;
} }
void Eliminate_Aux(const size_t node_index, const std::shared_ptr<Graph> graph, void Eliminate_Aux(const size_t node_index, const std::shared_ptr<Graph> &graph,
const std::shared_ptr<std::vector<std::vector<size_t>>> eli_list) { const std::shared_ptr<std::vector<std::vector<size_t>>> &eli_list) {
std::vector<size_t> eli; std::vector<size_t> eli;
eli.push_back(node_index); eli.push_back(node_index);
for (size_t i = 0; i < (size_t)graph->nodes[node_index].node_out.size(); i++) { for (size_t i = 0; i < (size_t)graph->nodes[node_index].node_out.size(); i++) {
@ -211,18 +211,18 @@ void Eliminate_Aux(const size_t node_index, const std::shared_ptr<Graph> graph,
} }
} }
std::shared_ptr<Graph> EliminateGraph(const std::shared_ptr<Graph> graph, std::shared_ptr<Graph> EliminateGraph(const std::shared_ptr<Graph> &graph,
const std::shared_ptr<std::vector<std::vector<size_t>>> eli_list, const std::shared_ptr<std::vector<std::vector<size_t>>> &eli_list,
const std::shared_ptr<std::vector<size_t>> index_list) { const std::shared_ptr<std::vector<size_t>> &index_list) {
MS_EXCEPTION_IF_NULL(graph); MS_EXCEPTION_IF_NULL(graph);
const std::set<OperatorType> type_list = { static const std::set<OperatorType> elementwise_type = {
OperatorType::kRecOneHot, OperatorType::kRecReLU, OperatorType::kRecLog, OperatorType::kRecExp, OperatorType::kRecReLU, OperatorType::kRecLog, OperatorType::kRecExp, OperatorType::kRecAdd,
OperatorType::kRecAdd, OperatorType::kRecElmWiseOp, OperatorType::kRecBiasAdd, OperatorType::kRecSub, OperatorType::kRecElmWiseOp, OperatorType::kRecBiasAdd, OperatorType::kRecSub, OperatorType::kRecMul,
OperatorType::kRecMul, OperatorType::kRecDiv, OperatorType::kRecSqueeze, OperatorType::kRecReduce, OperatorType::kRecDiv, OperatorType::kRecSqueeze, OperatorType::kRecReduce, OperatorType::kRecCast,
OperatorType::kRecCast, OperatorType::kRecReshape, OperatorType::kRecGatherV2}; OperatorType::kRecReshape, OperatorType::kRecGatherV2, OperatorType::kRecArgWithValue};
for (size_t node_index = 0; node_index < (size_t)graph->nodes.size(); node_index++) { for (size_t node_index = 0; node_index < (size_t)graph->nodes.size(); node_index++) {
auto type = graph->nodes[node_index].apply.op_type; auto type = graph->nodes[node_index].apply.op_type;
if (type_list.find(type) != type_list.end()) { if (elementwise_type.find(type) != elementwise_type.end()) {
Eliminate_Aux(node_index, graph, eli_list); Eliminate_Aux(node_index, graph, eli_list);
} }
} }
@ -250,12 +250,22 @@ std::shared_ptr<Graph> EliminateGraph(const std::shared_ptr<Graph> graph,
new_graph->nodes.push_back(graph->nodes[i]); new_graph->nodes.push_back(graph->nodes[i]);
auto *node_in = &new_graph->nodes[index_list->at(i)].node_in; auto *node_in = &new_graph->nodes[index_list->at(i)].node_in;
for (size_t j = 0; j < node_in->size(); j++) { for (size_t j = node_in->size(); j > 0; j--) {
node_in->at(j) = index_list->at(node_in->at(j)); bool IsEliminated = (index_list->at(node_in->at(j - 1)) == SIZE_MAX);
if (IsEliminated) {
node_in->erase(node_in->begin() + j - 1);
} else {
node_in->at(j - 1) = index_list->at(node_in->at(j - 1));
}
} }
auto *node_out = &new_graph->nodes[index_list->at(i)].node_out; auto *node_out = &new_graph->nodes[index_list->at(i)].node_out;
for (size_t j = 0; j < node_out->size(); j++) { for (size_t j = node_out->size(); j > 0; j--) {
node_out->at(j) = index_list->at(node_out->at(j)); bool IsEliminated = (index_list->at(node_out->at(j - 1)) == SIZE_MAX);
if (IsEliminated) {
node_out->erase(node_out->begin() + j - 1);
} else {
node_out->at(j - 1) = index_list->at(node_out->at(j - 1));
}
} }
} }
return new_graph; return new_graph;

19
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_parse_graph.h
View File

@ -47,6 +47,8 @@ const std::map<std::string, OperatorType> DictOpType{
{REDUCE_MIN, OperatorType::kRecReduce}, {REDUCE_MIN, OperatorType::kRecReduce},
{REDUCE_MEAN, OperatorType::kRecReduce}, {REDUCE_MEAN, OperatorType::kRecReduce},
{GATHERV2, OperatorType::kRecGatherV2}, {GATHERV2, OperatorType::kRecGatherV2},
{ARGMAXWITHVALUE, OperatorType::kRecArgWithValue},
{ARGMINWITHVALUE, OperatorType::kRecArgWithValue},
{RELU, OperatorType::kRecReLU}, {RELU, OperatorType::kRecReLU},
{"ReLU6", OperatorType::kRecReLU}, {"ReLU6", OperatorType::kRecReLU},
@ -59,6 +61,7 @@ const std::map<std::string, OperatorType> DictOpType{
{PRELU, OperatorType::kRecPReLU}, {PRELU, OperatorType::kRecPReLU},
{TRANSPOSE, OperatorType::kRecElmWiseOp},
{L2_NORMALIZE, OperatorType::kRecElmWiseOp}, {L2_NORMALIZE, OperatorType::kRecElmWiseOp},
{TENSOR_ADD, OperatorType::kRecElmWiseOp}, {TENSOR_ADD, OperatorType::kRecElmWiseOp},
{SUB, OperatorType::kRecElmWiseOp}, {SUB, OperatorType::kRecElmWiseOp},
@ -67,7 +70,7 @@ const std::map<std::string, OperatorType> DictOpType{
{REAL_DIV, OperatorType::kRecElmWiseOp}, {REAL_DIV, OperatorType::kRecElmWiseOp},
{SOFTMAX, OperatorType::kRecSoftmax}, {SOFTMAX, OperatorType::kRecSoftmax},
{LOG_SOFTMAX, OperatorType::kRecSoftmax}, {LOG_SOFTMAX, OperatorType::kRecSoftmax},
{SOFTMAX_CROSS_ENTROPY_WITH_LOGITS, OperatorType::kRecSoftmax}, {SOFTMAX_CROSS_ENTROPY_WITH_LOGITS, OperatorType::kRecSoftmaxCrossEntropyWithLogits},
{SQRT, OperatorType::kRecElmWiseOp}, {SQRT, OperatorType::kRecElmWiseOp},
{NEG, OperatorType::kRecElmWiseOp}, {NEG, OperatorType::kRecElmWiseOp},
{POW, OperatorType::kRecElmWiseOp}, {POW, OperatorType::kRecElmWiseOp},
@ -107,7 +110,7 @@ const std::map<std::string, OperatorType> DictOpType{
const TensorParam MakeTensor(int n, int c, int h, int w); const TensorParam MakeTensor(int n, int c, int h, int w);
Graph::NodeType MakeNewOperator(std::vector<std::shared_ptr<OperatorInfo>> ops, size_t iter_ops); Graph::NodeType MakeNewOperator(const std::vector<std::shared_ptr<OperatorInfo>> &ops, size_t iter_ops);
OperatorRec CompleteOperatorInputs(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops, OperatorRec CompleteOperatorInputs(const std::vector<std::shared_ptr<OperatorInfo>> &ops, const size_t iter_ops,
Graph::NodeType NewTensor); Graph::NodeType NewTensor);
@ -118,17 +121,17 @@ TensorParam Complete2DInputs(const std::vector<std::shared_ptr<OperatorInfo>> &o
std::shared_ptr<Graph> ParseGraph(const std::vector<std::shared_ptr<OperatorInfo>> &ops, std::shared_ptr<Graph> ParseGraph(const std::vector<std::shared_ptr<OperatorInfo>> &ops,
const std::vector<std::vector<std::string>> &input_tensor_names); const std::vector<std::vector<std::string>> &input_tensor_names);
void MakeEdge(const std::vector<std::vector<std::string>> &input_tensor_names, std::shared_ptr<Graph> graph); void MakeEdge(const std::vector<std::vector<std::string>> &input_tensor_names, const std::shared_ptr<Graph> &graph);
size_t GetIndexInInputTensorNames(const std::vector<std::vector<std::string>> &input_tensor_names, size_t GetIndexInInputTensorNames(const std::vector<std::vector<std::string>> &input_tensor_names,
const std::string &input_name); const std::string &input_name);
void Eliminate_Aux(const size_t node_index, const std::shared_ptr<Graph> graph, void Eliminate_Aux(const size_t node_index, const std::shared_ptr<Graph> &graph,
const std::shared_ptr<std::vector<std::vector<size_t>>> eli_list); const std::shared_ptr<std::vector<std::vector<size_t>>> &eli_list);
std::shared_ptr<Graph> EliminateGraph(const std::shared_ptr<Graph> graph, std::shared_ptr<Graph> EliminateGraph(const std::shared_ptr<Graph> &graph,
const std::shared_ptr<std::vector<std::vector<size_t>>> eli_list, const std::shared_ptr<std::vector<std::vector<size_t>>> &eli_list,
const std::shared_ptr<std::vector<size_t>> index_list); const std::shared_ptr<std::vector<size_t>> &index_list);
} // namespace parallel } // namespace parallel
} // namespace mindspore } // namespace mindspore
#endif // PARALLEL_AUTO_PARALLEL_REC_PARSE_GRAPH_H_ #endif // PARALLEL_AUTO_PARALLEL_REC_PARSE_GRAPH_H_

55
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_partition.cc
View File

@ -68,19 +68,24 @@ double GetWeights(const Graph::NodeType &node) {
auto cost_ptr = std::make_shared<CostBiasAdd>(); auto cost_ptr = std::make_shared<CostBiasAdd>();
return cost_ptr->GetMinCostIn(); return cost_ptr->GetMinCostIn();
} else if (op.op_type == OperatorType::kRecOneHot || op.op_type == OperatorType::kRecLog || } else if (op.op_type == OperatorType::kRecLog || op.op_type == OperatorType::kRecExp ||
op.op_type == OperatorType::kRecExp || op.op_type == OperatorType::kRecAdd || op.op_type == OperatorType::kRecAdd || op.op_type == OperatorType::kRecSub ||
op.op_type == OperatorType::kRecSub || op.op_type == OperatorType::kRecMul || op.op_type == OperatorType::kRecMul || op.op_type == OperatorType::kRecDiv ||
op.op_type == OperatorType::kRecDiv || op.op_type == OperatorType::kRecSqueeze || op.op_type == OperatorType::kRecSqueeze || op.op_type == OperatorType::kRecCast) {
op.op_type == OperatorType::kRecCast) {
// For element-wise op // For element-wise op
auto cost_ptr = std::make_shared<CostCommon>(); auto cost_ptr = std::make_shared<CostCommon>();
return cost_ptr->GetMinCostIn(); return cost_ptr->GetMinCostIn();
} else if (op.op_type == OperatorType::kRecUnkownType || op.op_type == OperatorType::kRecPReLU || } else if (op.op_type == OperatorType::kRecBatchNorm || op.op_type == OperatorType::kRecOneHot ||
op.op_type == OperatorType::kRecBatchNorm || op.op_type == OperatorType::kRecSoftmax || op.op_type == OperatorType::kRecPReLU || op.op_type == OperatorType::kRecSoftmax ||
op.op_type == OperatorType::kRecSparseSoftmaxCrossEntropyWithLogits) { op.op_type == OperatorType::kRecSparseSoftmaxCrossEntropyWithLogits ||
// For unprocessed type op.op_type == OperatorType::kRecSoftmaxCrossEntropyWithLogits) {
// For BatchParallel op
auto cost_ptr = std::make_shared<CostBatchParallel>();
return cost_ptr->GetMaxCostIn();
} else if (op.op_type == OperatorType::kRecUnkownType) {
// For Unkown type
return 0.0; return 0.0;
} else { } else {
MS_LOG(EXCEPTION) << "Failure: GetOperatorWeight failed."; MS_LOG(EXCEPTION) << "Failure: GetOperatorWeight failed.";
@ -88,7 +93,7 @@ double GetWeights(const Graph::NodeType &node) {
} }
// Sort all the nodes by their weights // Sort all the nodes by their weights
std::vector<size_t> SortByWeight(const std::shared_ptr<Graph> graph) { std::vector<size_t> SortByWeight(const std::shared_ptr<Graph> &graph) {
MS_EXCEPTION_IF_NULL(graph); MS_EXCEPTION_IF_NULL(graph);
std::vector<std::pair<double, size_t>> weight_to_node_index; std::vector<std::pair<double, size_t>> weight_to_node_index;
@ -119,7 +124,7 @@ std::vector<size_t> SortByWeight(const std::shared_ptr<Graph> graph) {
// Get optimal strategy to partition the target node // Get optimal strategy to partition the target node
StrategyRec PartitionNode(const Graph::NodeType &node, StrategyRec PartitionNode(const Graph::NodeType &node,
const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy, const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy,
std::shared_ptr<Graph> graph) { const std::shared_ptr<Graph> &graph) {
bool enable_conv_chw_partition = false; bool enable_conv_chw_partition = false;
MS_EXCEPTION_IF_NULL(graph); MS_EXCEPTION_IF_NULL(graph);
@ -158,19 +163,26 @@ StrategyRec PartitionNode(const Graph::NodeType &node,
auto cost_ptr = std::make_shared<CostBiasAdd>(); auto cost_ptr = std::make_shared<CostBiasAdd>();
return cost_ptr->GetOptimalStr(node, node_name_to_strategy, *graph); return cost_ptr->GetOptimalStr(node, node_name_to_strategy, *graph);
} else if (node.apply.op_type == OperatorType::kRecOneHot || node.apply.op_type == OperatorType::kRecLog || } else if (node.apply.op_type == OperatorType::kRecLog || node.apply.op_type == OperatorType::kRecExp ||
node.apply.op_type == OperatorType::kRecExp || node.apply.op_type == OperatorType::kRecAdd || node.apply.op_type == OperatorType::kRecAdd || node.apply.op_type == OperatorType::kRecSub ||
node.apply.op_type == OperatorType::kRecSub || node.apply.op_type == OperatorType::kRecMul || node.apply.op_type == OperatorType::kRecMul || node.apply.op_type == OperatorType::kRecDiv ||
node.apply.op_type == OperatorType::kRecDiv || node.apply.op_type == OperatorType::kRecSqueeze || node.apply.op_type == OperatorType::kRecSqueeze || node.apply.op_type == OperatorType::kRecCast) {
node.apply.op_type == OperatorType::kRecCast) {
// For element-wise op // For element-wise op
auto cost_ptr = std::make_shared<CostCommon>(); auto cost_ptr = std::make_shared<CostCommon>();
return cost_ptr->GetOptimalStr(node, node_name_to_strategy, *graph); return cost_ptr->GetOptimalStr(node, node_name_to_strategy, *graph);
} else if (node.apply.op_type == OperatorType::kRecUnkownType || node.apply.op_type == OperatorType::kRecPReLU || } else if (node.apply.op_type == OperatorType::kRecBatchNorm || node.apply.op_type == OperatorType::kRecOneHot ||
node.apply.op_type == OperatorType::kRecBatchNorm || node.apply.op_type == OperatorType::kRecSoftmax || node.apply.op_type == OperatorType::kRecPReLU || node.apply.op_type == kRecSoftmax ||
node.apply.op_type == OperatorType::kRecSparseSoftmaxCrossEntropyWithLogits) { node.apply.op_type == OperatorType::kRecSparseSoftmaxCrossEntropyWithLogits) {
// For unprocessed type // For BatchParallel type
auto cost_ptr = std::make_shared<CostBatchParallel>();
return cost_ptr->GetOptimalStr(node);
} else if (node.apply.op_type == OperatorType::kRecSoftmaxCrossEntropyWithLogits) {
// For SoftmaxCrossEntropyWithLogits type
auto cost_ptr = std::make_shared<CostSoftmaxCrossEntropyWithLogits>();
return cost_ptr->GetOptimalStr(node);
} else if (node.apply.op_type == OperatorType::kRecUnkownType) {
// For Unkown type
StrategyRec default_strategy; StrategyRec default_strategy;
return default_strategy; return default_strategy;
} else { } else {
@ -179,7 +191,8 @@ StrategyRec PartitionNode(const Graph::NodeType &node,
} }
// Parttion graph into all devices. // Parttion graph into all devices.
Status PartitionForAllDevices(const size_t num_device, const double device_memory, std::shared_ptr<Graph> graph) { Status PartitionForAllDevices(const size_t num_device, const double device_memory,
const std::shared_ptr<Graph> &graph) {
if (num_device < 1) { if (num_device < 1) {
MS_LOG(EXCEPTION) << "ERROR: Number of devices can't be " << num_device << "."; MS_LOG(EXCEPTION) << "ERROR: Number of devices can't be " << num_device << ".";
} }
@ -249,7 +262,7 @@ Graph::NodeType ApplyStrToTensor(Graph::NodeType Node) {
return Node; return Node;
} }
Status DevicesMemoryControl(const size_t num_device, const double device_memory, std::shared_ptr<Graph> graph) { Status DevicesMemoryControl(const size_t num_device, const double device_memory, const std::shared_ptr<Graph> &graph) {
MS_EXCEPTION_IF_NULL(graph); MS_EXCEPTION_IF_NULL(graph);
if (num_device == 0) { if (num_device == 0) {
MS_LOG(EXCEPTION) << "Failure: device number is 0."; MS_LOG(EXCEPTION) << "Failure: device number is 0.";

8
mindspore/ccsrc/parallel/auto_parallel/rec_core/rec_partition.h
View File

@ -32,19 +32,19 @@
namespace mindspore { namespace mindspore {
namespace parallel { namespace parallel {
std::vector<size_t> SortByWeight(const std::shared_ptr<Graph> graph); std::vector<size_t> SortByWeight(const std::shared_ptr<Graph> &graph);
double GetWeights(const Graph::NodeType &node); double GetWeights(const Graph::NodeType &node);
StrategyRec PartitionNode(const Graph::NodeType &node, StrategyRec PartitionNode(const Graph::NodeType &node,
const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy, const std::vector<std::pair<std::string, StrategyRec>> &node_name_to_strategy,
std::shared_ptr<Graph> graph); const std::shared_ptr<Graph> &graph);
Status PartitionForAllDevices(const size_t num_device, const double device_memory, std::shared_ptr<Graph> graph); Status PartitionForAllDevices(const size_t num_device, const double device_memory, const std::shared_ptr<Graph> &graph);
Graph::NodeType ApplyStrToTensor(Graph::NodeType Node); Graph::NodeType ApplyStrToTensor(Graph::NodeType Node);
Status DevicesMemoryControl(const size_t num_device, const double device_memory, std::shared_ptr<Graph> graph); Status DevicesMemoryControl(const size_t num_device, const double device_memory, const std::shared_ptr<Graph> &graph);
size_t GetDataTypeSize(const TensorType &type); size_t GetDataTypeSize(const TensorType &type);
} // namespace parallel } // namespace parallel