@@ -85,8 +85,9 @@ static Status CheckEngineTypeSupport(const NodePtr &node, OpEngineType engine_ty | |||||
} else { | } else { | ||||
ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"}, | ErrorManager::GetInstance().ATCReportErrMessage("E14001", {"opname", "optype", "value", "reason"}, | ||||
{op_desc->GetName(), op_desc->GetType(), "engine type", | {op_desc->GetName(), op_desc->GetType(), "engine type", | ||||
"it only support kEngineNameDefault/kAIcoreEngine/kVectorEngine"}); | |||||
GELOGE(FAILED, "CheckEngineType: engine type: %d not support.", static_cast<int>(engine_type)); | |||||
"it only support default/AIcoreEngine/VectorEngine"}); | |||||
GELOGE(FAILED, "[Check][EngineType]value:%d not support, " | |||||
"only support default/AIcoreEngine/VectorEngine now", static_cast<int>(engine_type)); | |||||
return FAILED; | return FAILED; | ||||
} | } | ||||
@@ -190,17 +191,20 @@ static Status AddInputs(const ComputeGraphPtr &graph, const NodePtr &node, const | |||||
(void)AttrUtils::SetBool(data_op, "_is_single_op", true); | (void)AttrUtils::SetBool(data_op, "_is_single_op", true); | ||||
GE_CHK_BOOL_EXEC(data_op->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add input desc fail"); | |||||
GE_CHK_BOOL_EXEC(data_op->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add output desc fail"); | |||||
GE_CHK_BOOL_EXEC(data_op->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, | |||||
"[Add][InputDesc]fail for node:%s", data_op->GetName().c_str()); | |||||
GE_CHK_BOOL_EXEC(data_op->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, | |||||
"[Add][OutputDesc]fail for node:%s", data_op->GetName().c_str()); | |||||
if (attr) { | if (attr) { | ||||
GE_CHK_BOOL_EXEC(AttrUtils::SetInt(data_op, ATTR_NAME_INDEX, index), return FAILED, "Set index fail"); | |||||
GE_CHK_BOOL_EXEC(AttrUtils::SetInt(data_op, ATTR_NAME_INDEX, index), return FAILED, | |||||
"[Set][Attr:%s]fail for node:%s", ATTR_NAME_INDEX.c_str(), data_op->GetName().c_str()); | |||||
} | } | ||||
ge::NodePtr arg_node = graph->AddNode(data_op); | ge::NodePtr arg_node = graph->AddNode(data_op); | ||||
GE_CHK_BOOL_EXEC(arg_node != nullptr, return FAILED, "Insert Data node fail"); | GE_CHK_BOOL_EXEC(arg_node != nullptr, return FAILED, "Insert Data node fail"); | ||||
GE_CHK_STATUS(GraphUtils::AddEdge(arg_node->GetOutDataAnchor(0), node->GetInDataAnchor(index)), | GE_CHK_STATUS(GraphUtils::AddEdge(arg_node->GetOutDataAnchor(0), node->GetInDataAnchor(index)), | ||||
"Add edge[%s->%s] fail", data_op->GetName().c_str(), node->GetName().c_str()); | |||||
"[Add][Edge]fail from node:%s to node:%s", data_op->GetName().c_str(), node->GetName().c_str()); | |||||
return SUCCESS; | return SUCCESS; | ||||
} | } | ||||
@@ -215,20 +219,23 @@ static Status AddOutputs(const ComputeGraphPtr &graph, const NodePtr &node, cons | |||||
for (const auto &out_desc : outputs) { | for (const auto &out_desc : outputs) { | ||||
GeTensorDesc tensor = out_desc.GetTensorDesc(); | GeTensorDesc tensor = out_desc.GetTensorDesc(); | ||||
TensorUtils::SetInputTensor(tensor, true); | TensorUtils::SetInputTensor(tensor, true); | ||||
GE_CHK_BOOL_EXEC(op_desc->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add input desc fail."); | |||||
GE_CHK_BOOL_EXEC(op_desc->AddInputDesc(tensor) == GRAPH_SUCCESS, return FAILED, | |||||
"[Add][InputDesc]fail for node:%s", op_desc->GetName().c_str()); | |||||
TensorUtils::SetInputTensor(tensor, false); | TensorUtils::SetInputTensor(tensor, false); | ||||
TensorUtils::SetOutputTensor(tensor, true); | TensorUtils::SetOutputTensor(tensor, true); | ||||
GE_CHK_BOOL_EXEC(op_desc->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, "Add output desc fail."); | |||||
GE_CHK_BOOL_EXEC(op_desc->AddOutputDesc(tensor) == GRAPH_SUCCESS, return FAILED, | |||||
"[Add][OutputDesc]fail for node:%s", op_desc->GetName().c_str()); | |||||
count++; | count++; | ||||
} | } | ||||
GE_CHECK_NOTNULL_EXEC(graph, return PARAM_INVALID); | GE_CHECK_NOTNULL_EXEC(graph, return PARAM_INVALID); | ||||
ge::NodePtr out_node = graph->AddNode(op_desc); | ge::NodePtr out_node = graph->AddNode(op_desc); | ||||
GE_CHK_BOOL_EXEC(out_node != nullptr, return FAILED, "Insert Output node fail"); | |||||
GE_CHK_BOOL_EXEC(out_node != nullptr, return FAILED, | |||||
"[Add][Node:%s]fail in graph:%u", op_desc->GetName().c_str(), graph->GetGraphID()); | |||||
GE_CHECK_NOTNULL_EXEC(node, return PARAM_INVALID); | GE_CHECK_NOTNULL_EXEC(node, return PARAM_INVALID); | ||||
for (int32_t i = 0; i < count; ++i) { | for (int32_t i = 0; i < count; ++i) { | ||||
GE_CHK_STATUS(GraphUtils::AddEdge(node->GetOutDataAnchor(i), out_node->GetInDataAnchor(i)), | GE_CHK_STATUS(GraphUtils::AddEdge(node->GetOutDataAnchor(i), out_node->GetInDataAnchor(i)), | ||||
"Add edge[%s->%s] fail", node->GetName().c_str(), out_node->GetName().c_str()); | |||||
"[Add][Edge]fail from node:%s to node:%s", node->GetName().c_str(), out_node->GetName().c_str()); | |||||
} | } | ||||
return SUCCESS; | return SUCCESS; | ||||
@@ -248,7 +255,7 @@ static void GetOpsProtoPath(string &opsproto_path) { | |||||
return; | return; | ||||
} | } | ||||
string path_base = PluginManager::GetPath(); | string path_base = PluginManager::GetPath(); | ||||
GELOGI("path_base is %s.", path_base.c_str()); | |||||
GELOGI("path_base is %s", path_base.c_str()); | |||||
path_base = path_base.substr(0, path_base.rfind('/')); | path_base = path_base.substr(0, path_base.rfind('/')); | ||||
path_base = path_base.substr(0, path_base.rfind('/') + 1); | path_base = path_base.substr(0, path_base.rfind('/') + 1); | ||||
opsproto_path = (path_base + "ops/op_proto/custom/" + ":") + (path_base + "ops/op_proto/built-in/"); | opsproto_path = (path_base + "ops/op_proto/custom/" + ":") + (path_base + "ops/op_proto/built-in/"); | ||||
@@ -333,7 +340,7 @@ Status GeGenerator::Initialize(const map<string, string> &options, OmgContext &o | |||||
ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOpsProtoInit); | ErrorManager::GetInstance().SetStage(ErrorMessage::kInitialize, ErrorMessage::kOpsProtoInit); | ||||
string opsproto_path; | string opsproto_path; | ||||
GetOpsProtoPath(opsproto_path); | GetOpsProtoPath(opsproto_path); | ||||
GELOGI("Get opsproto path is %s.", opsproto_path.c_str()); | |||||
GELOGI("Get opsproto path is %s", opsproto_path.c_str()); | |||||
OpsProtoManager *manager = OpsProtoManager::Instance(); | OpsProtoManager *manager = OpsProtoManager::Instance(); | ||||
map<string, string> option_tmp; | map<string, string> option_tmp; | ||||
option_tmp.emplace(std::pair<string, string>(string("ge.opsProtoLibPath"), opsproto_path)); | option_tmp.emplace(std::pair<string, string>(string("ge.opsProtoLibPath"), opsproto_path)); | ||||
@@ -712,7 +719,7 @@ Status GeGenerator::BuildSingleOp(OpDescPtr &op_desc, const vector<GeTensor> &in | |||||
auto node = comp_graph->FindNode(op_desc->GetName()); | auto node = comp_graph->FindNode(op_desc->GetName()); | ||||
Status ret = CheckEngineTypeSupport(node, engine_type); | Status ret = CheckEngineTypeSupport(node, engine_type); | ||||
if (ret != SUCCESS) { | if (ret != SUCCESS) { | ||||
GELOGE(ret, "check engine type failed"); | |||||
GELOGE(ret, "[Check][EngineType]value:%d for node:%s not support", engine_type, node->GetName().c_str()); | |||||
return ret; | return ret; | ||||
} | } | ||||
} | } | ||||
@@ -786,9 +793,9 @@ Status GeGenerator::BuildSingleOpModel(OpDescPtr &op_desc, const vector<GeTensor | |||||
const vector<GeTensor> &outputs, OpEngineType engine_type, | const vector<GeTensor> &outputs, OpEngineType engine_type, | ||||
ModelBufferData &model_buff) { | ModelBufferData &model_buff) { | ||||
ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kOther); | ErrorManager::GetInstance().SetStage(ErrorMessage::kModelCompile, ErrorMessage::kOther); | ||||
GELOGI("Start to build single op online, input size: %zu, output size: %zu.", inputs.size(), outputs.size()); | |||||
GELOGI("Start to build single op online, input size: %zu, output size: %zu", inputs.size(), outputs.size()); | |||||
Status status = BuildSingleOp(op_desc, inputs, outputs, kFileNameSuffix, engine_type, model_buff, false); | Status status = BuildSingleOp(op_desc, inputs, outputs, kFileNameSuffix, engine_type, model_buff, false); | ||||
GELOGI("Finish build single online model, status: %u.", status); | |||||
GELOGI("Finish build single online model, status: %u", status); | |||||
return status; | return status; | ||||
} | } | ||||
@@ -33,13 +33,21 @@ using std::queue; | |||||
namespace ge { | namespace ge { | ||||
LogicalStreamPass::LogicalStreamPass(const string &name) : name_(name) {} | LogicalStreamPass::LogicalStreamPass(const string &name) : name_(name) {} | ||||
const string &LogicalStreamPass::GetName() const { return name_; } | |||||
const string &LogicalStreamPass::GetName() const { | |||||
return name_; | |||||
} | |||||
bool LogicalStreamPass::IsEngineSkip(const Subgraph &subgraph) const { return subgraph.engine_conf.skip_assign_stream; } | |||||
bool LogicalStreamPass::IsEngineSkip(const Subgraph &subgraph) const { | |||||
return subgraph.engine_conf.skip_assign_stream; | |||||
} | |||||
bool LogicalStreamPass::IsEngineAttach(const Subgraph &subgraph) const { return subgraph.engine_conf.attach; } | |||||
bool LogicalStreamPass::IsEngineAttach(const Subgraph &subgraph) const { | |||||
return subgraph.engine_conf.attach; | |||||
} | |||||
bool LogicalStreamPass::IsEngineIndependent(const Subgraph &subgraph) const { return subgraph.engine_conf.independent; } | |||||
bool LogicalStreamPass::IsEngineIndependent(const Subgraph &subgraph) const { | |||||
return subgraph.engine_conf.independent; | |||||
} | |||||
bool LogicalStreamPass::HasStreamLabel(const Subgraph &subgraph) const { | bool LogicalStreamPass::HasStreamLabel(const Subgraph &subgraph) const { | ||||
return !subgraph.subgraph_info.GetStreamLabel().empty(); | return !subgraph.subgraph_info.GetStreamLabel().empty(); | ||||
@@ -60,14 +68,14 @@ Status AssignByLabelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr> & | |||||
// Subgraphs of the same stream_label are assigned to the same stream, | // Subgraphs of the same stream_label are assigned to the same stream, | ||||
// and different stream_labels are assigned new streams. | // and different stream_labels are assigned new streams. | ||||
auto iter = label_streams.find(stream_label); | auto iter = label_streams.find(stream_label); | ||||
if (iter != label_streams.end()) { | |||||
subgraph->stream_id = iter->second; | |||||
} else { | |||||
if (iter == label_streams.end()) { | |||||
subgraph->stream_id = next_stream; | subgraph->stream_id = next_stream; | ||||
GELOGI("Assign new stream %ld for label %s", next_stream, stream_label.c_str()); | |||||
GELOGI("Assign new stream %ld for label %s.", next_stream, stream_label.c_str()); | |||||
label_streams.emplace(stream_label, next_stream); | label_streams.emplace(stream_label, next_stream); | ||||
++next_stream; | |||||
next_stream++; | |||||
} else { | |||||
subgraph->stream_id = iter->second; | |||||
} | } | ||||
changed = true; | changed = true; | ||||
} | } | ||||
@@ -92,15 +100,15 @@ Status IndependentStreamPass::Run(ComputeGraphPtr graph, const vector<SubgraphPt | |||||
const string &stream_label = subgraph->subgraph_info.GetStreamLabel(); | const string &stream_label = subgraph->subgraph_info.GetStreamLabel(); | ||||
auto &label_streams = engine_streams[engine]; | auto &label_streams = engine_streams[engine]; | ||||
auto iter = label_streams.find(stream_label); | auto iter = label_streams.find(stream_label); | ||||
if (iter != label_streams.end()) { | |||||
subgraph->stream_id = iter->second; | |||||
} else { | |||||
if (iter == label_streams.end()) { | |||||
subgraph->stream_id = next_stream; | subgraph->stream_id = next_stream; | ||||
GELOGI("Assign new independent stream %ld for engine %s (label: %s)", next_stream, engine.c_str(), | |||||
GELOGI("Assign new independent stream %ld for engine %s (label: %s).", next_stream, engine.c_str(), | |||||
stream_label.c_str()); | stream_label.c_str()); | ||||
label_streams.emplace(stream_label, next_stream); | label_streams.emplace(stream_label, next_stream); | ||||
++next_stream; | |||||
next_stream++; | |||||
} else { | |||||
subgraph->stream_id = iter->second; | |||||
} | } | ||||
changed = true; | changed = true; | ||||
} | } | ||||
@@ -121,13 +129,15 @@ Status AssignByDependencyPass::Run(ComputeGraphPtr graph, const vector<SubgraphP | |||||
} | } | ||||
SubgraphPtr reusable_subgraph = GetReusableSubgraph(subgraph, end_subgraph_map, pld_subgraph_map); | SubgraphPtr reusable_subgraph = GetReusableSubgraph(subgraph, end_subgraph_map, pld_subgraph_map); | ||||
if (reusable_subgraph != nullptr) { | |||||
if (reusable_subgraph == nullptr) { | |||||
(void)AssignNewStream(subgraph); | |||||
} else { | |||||
if (HasAssignedStream(*reusable_subgraph)) { | if (HasAssignedStream(*reusable_subgraph)) { | ||||
subgraph->stream_id = reusable_subgraph->stream_id; | subgraph->stream_id = reusable_subgraph->stream_id; | ||||
} else { | } else { | ||||
int64_t stream_id = AssignNewStream(reusable_subgraph); | int64_t stream_id = AssignNewStream(reusable_subgraph); | ||||
subgraph->stream_id = stream_id; | subgraph->stream_id = stream_id; | ||||
GELOGI("Reusable subgraph %s has not been assigned a stream, now assign new stream %ld", | |||||
GELOGI("Reusable subgraph %s has not been assigned a stream, now assign new stream %ld.", | |||||
reusable_subgraph->name.c_str(), stream_id); | reusable_subgraph->name.c_str(), stream_id); | ||||
} | } | ||||
@@ -137,11 +147,9 @@ Status AssignByDependencyPass::Run(ComputeGraphPtr graph, const vector<SubgraphP | |||||
subgraph->reused_subgraph = reusable_subgraph; | subgraph->reused_subgraph = reusable_subgraph; | ||||
reused_subgraphs_.emplace_back(subgraph, reusable_subgraph); | reused_subgraphs_.emplace_back(subgraph, reusable_subgraph); | ||||
GELOGI("Subgraph %s of engine %s reuses stream of subgraph %s of engine %s", subgraph->name.c_str(), | |||||
GELOGI("Subgraph %s of engine %s reuses stream of subgraph %s of engine %s.", subgraph->name.c_str(), | |||||
subgraph->engine_conf.id.c_str(), reusable_subgraph->name.c_str(), | subgraph->engine_conf.id.c_str(), reusable_subgraph->name.c_str(), | ||||
reusable_subgraph->engine_conf.id.c_str()); | reusable_subgraph->engine_conf.id.c_str()); | ||||
} else { | |||||
(void)AssignNewStream(subgraph); | |||||
} | } | ||||
changed = true; | changed = true; | ||||
} | } | ||||
@@ -191,13 +199,15 @@ bool AssignByDependencyPass::CouldReuse(const SubgraphPtr &subgraph, const Subgr | |||||
auto iter = pld_subgraph_map.find(end_pld_pair.second); | auto iter = pld_subgraph_map.find(end_pld_pair.second); | ||||
if (iter != pld_subgraph_map.end()) { | if (iter != pld_subgraph_map.end()) { | ||||
const SubgraphPtr &pred_subgraph_succ = iter->second; | const SubgraphPtr &pred_subgraph_succ = iter->second; | ||||
if (pred_subgraph_succ != subgraph && pred_subgraph_succ->engine_conf.id == pred_subgraph->engine_conf.id) { | |||||
if ((pred_subgraph_succ != subgraph) && | |||||
(pred_subgraph_succ->engine_conf.id == pred_subgraph->engine_conf.id)) { | |||||
return false; | return false; | ||||
} | } | ||||
} | } | ||||
} | } | ||||
if ((subgraph->engine_conf.id == pred_subgraph->engine_conf.id) || IsEngineAttach(*subgraph)) { | |||||
if ((subgraph->engine_conf.id == pred_subgraph->engine_conf.id) || | |||||
IsEngineAttach(*subgraph)) { | |||||
return true; | return true; | ||||
} | } | ||||
@@ -249,7 +259,7 @@ int64_t AssignByDependencyPass::AssignNewStream(SubgraphPtr subgraph) { | |||||
engine_stream_num_[engine_name] = stream_id + 1; | engine_stream_num_[engine_name] = stream_id + 1; | ||||
} | } | ||||
GELOGI("Subgraph %s assigns new temp stream %ld (engine: %s)", subgraph->name.c_str(), stream_id, | |||||
GELOGI("Subgraph %s assigns new temp stream %ld (engine: %s).", subgraph->name.c_str(), stream_id, | |||||
engine_name.c_str()); | engine_name.c_str()); | ||||
return stream_id; | return stream_id; | ||||
@@ -282,7 +292,7 @@ void AssignByDependencyPass::UpdateAssignedSubgraphs(Context &context) { | |||||
GELOGI("Subgraph %s of engine %s reuses default stream %ld.", subgraph->name.c_str(), | GELOGI("Subgraph %s of engine %s reuses default stream %ld.", subgraph->name.c_str(), | ||||
subgraph->engine_conf.id.c_str(), context.default_stream); | subgraph->engine_conf.id.c_str(), context.default_stream); | ||||
} else { | } else { | ||||
GELOGI("Stream of subgraph %s has been updated to %ld", subgraph->name.c_str(), subgraph->stream_id); | |||||
GELOGI("Stream of subgraph %s has been updated to %ld.", subgraph->name.c_str(), subgraph->stream_id); | |||||
} | } | ||||
} | } | ||||
} | } | ||||
@@ -293,7 +303,7 @@ void AssignByDependencyPass::UpdateReusedSubgraphs() { | |||||
auto &cur_subgraph = item.first; | auto &cur_subgraph = item.first; | ||||
auto &reused_graph = item.second; | auto &reused_graph = item.second; | ||||
cur_subgraph->stream_id = reused_graph->stream_id; | cur_subgraph->stream_id = reused_graph->stream_id; | ||||
GELOGI("Stream of subgraph %s has been updated to %ld", cur_subgraph->name.c_str(), cur_subgraph->stream_id); | |||||
GELOGI("Stream of subgraph %s has been updated to %ld.", cur_subgraph->name.c_str(), cur_subgraph->stream_id); | |||||
} | } | ||||
} | } | ||||
@@ -330,7 +340,7 @@ Status NodeStreamUpdatePass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr | |||||
engine_name.c_str()); | engine_name.c_str()); | ||||
return INTERNAL_ERROR; | return INTERNAL_ERROR; | ||||
} else { | } else { | ||||
GELOGI("Subgraph %s is assigned stream %ld (engine: %s)", subgraph->name.c_str(), subgraph->stream_id, | |||||
GELOGI("Subgraph %s is assigned stream %ld (engine: %s).", subgraph->name.c_str(), subgraph->stream_id, | |||||
engine_name.c_str()); | engine_name.c_str()); | ||||
} | } | ||||
} | } | ||||
@@ -353,11 +363,11 @@ Status NodeStreamUpdatePass::Run(ComputeGraphPtr graph, const vector<SubgraphPtr | |||||
GELOGD("Node %s of type %s in subgraph %s is assigned parent stream %ld (engine: %s).", node->GetName().c_str(), | GELOGD("Node %s of type %s in subgraph %s is assigned parent stream %ld (engine: %s).", node->GetName().c_str(), | ||||
node->GetType().c_str(), subgraph->name.c_str(), context.default_stream, engine_name.c_str()); | node->GetType().c_str(), subgraph->name.c_str(), context.default_stream, engine_name.c_str()); | ||||
} else if (IsEngineSkip(*subgraph) && node->GetInNodes().empty()) { | } else if (IsEngineSkip(*subgraph) && node->GetInNodes().empty()) { | ||||
GELOGD("Node %s of type %s in subgraph %s doesn't need to assign a stream (engine: %s)", | |||||
GELOGD("Node %s of type %s in subgraph %s doesn't need to assign a stream (engine: %s).", | |||||
node->GetName().c_str(), node->GetType().c_str(), subgraph->name.c_str(), engine_name.c_str()); | node->GetName().c_str(), node->GetType().c_str(), subgraph->name.c_str(), engine_name.c_str()); | ||||
} else { | } else { | ||||
node->GetOpDesc()->SetStreamId(stream_id); | node->GetOpDesc()->SetStreamId(stream_id); | ||||
GELOGD("Node %s of type %s in subgraph %s is assigned stream %ld (engine: %s)", node->GetName().c_str(), | |||||
GELOGD("Node %s of type %s in subgraph %s is assigned stream %ld (engine: %s).", node->GetName().c_str(), | |||||
node->GetType().c_str(), subgraph->name.c_str(), stream_id, engine_name.c_str()); | node->GetType().c_str(), subgraph->name.c_str(), stream_id, engine_name.c_str()); | ||||
} | } | ||||
} | } | ||||
@@ -387,7 +397,7 @@ int64_t UpdateForSkippedEnginePass::GetSingleInoutStream(const NodePtr &node) co | |||||
if (stream_ids.size() == 1) { | if (stream_ids.size() == 1) { | ||||
int64_t stream_id = *(stream_ids.begin()); | int64_t stream_id = *(stream_ids.begin()); | ||||
GELOGI("The stream of all input and output nodes of node %s (type: %s) is %ld", node->GetName().c_str(), | |||||
GELOGI("The stream of all input and output nodes of node %s (type: %s) is %ld.", node->GetName().c_str(), | |||||
node->GetType().c_str(), stream_id); | node->GetType().c_str(), stream_id); | ||||
return stream_id; | return stream_id; | ||||
} | } | ||||
@@ -406,7 +416,7 @@ Status UpdateForSkippedEnginePass::Run(ComputeGraphPtr graph, const vector<Subgr | |||||
auto op_desc = node->GetOpDesc(); | auto op_desc = node->GetOpDesc(); | ||||
GE_CHECK_NOTNULL(op_desc); | GE_CHECK_NOTNULL(op_desc); | ||||
auto stream_id = op_desc->GetStreamId(); | auto stream_id = op_desc->GetStreamId(); | ||||
if (stream_id != kInvalidStream && !HasStreamLabel(*subgraph)) { | |||||
if ((stream_id != kInvalidStream) && !HasStreamLabel(*subgraph)) { | |||||
ops_without_label.emplace(op_desc); | ops_without_label.emplace(op_desc); | ||||
} | } | ||||
} | } | ||||
@@ -427,7 +437,7 @@ Status UpdateForSkippedEnginePass::Run(ComputeGraphPtr graph, const vector<Subgr | |||||
int64_t inout_stream = GetSingleInoutStream(node); | int64_t inout_stream = GetSingleInoutStream(node); | ||||
if (inout_stream != kInvalidStream) { | if (inout_stream != kInvalidStream) { | ||||
op_desc->SetStreamId(inout_stream); | op_desc->SetStreamId(inout_stream); | ||||
GELOGI("Node %s of type %s reassign to stream %ld from stream %ld", node->GetName().c_str(), | |||||
GELOGI("Node %s of type %s reassign to stream %ld from stream %ld.", node->GetName().c_str(), | |||||
node->GetType().c_str(), inout_stream, stream_id); | node->GetType().c_str(), inout_stream, stream_id); | ||||
} | } | ||||
} | } | ||||
@@ -455,7 +465,7 @@ Status AllReduceParallelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPt | |||||
return NOT_CHANGED; | return NOT_CHANGED; | ||||
} | } | ||||
GELOGI("AllReduceParallelPass is enabled"); | |||||
GELOGI("AllReduceParallelPass is enabled."); | |||||
GE_DUMP(graph, "BeforeAllReduceParallel"); | GE_DUMP(graph, "BeforeAllReduceParallel"); | ||||
// All successors of HcomAllReduce. | // All successors of HcomAllReduce. | ||||
@@ -463,7 +473,7 @@ Status AllReduceParallelPass::Run(ComputeGraphPtr graph, const vector<SubgraphPt | |||||
for (const NodePtr &node : graph->GetDirectNode()) { | for (const NodePtr &node : graph->GetDirectNode()) { | ||||
if (!IsHcomNode(node->GetType()) || | if (!IsHcomNode(node->GetType()) || | ||||
node->GetInDataNodes().size() <= 1) { | |||||
(node->GetInDataNodes().size() <= 1)) { | |||||
continue; | continue; | ||||
} | } | ||||
@@ -565,7 +575,7 @@ Status LogicalStreamAllocator::Assign(const ComputeGraphPtr &root_graph, const G | |||||
RefreshContinuousStreams(root_graph); | RefreshContinuousStreams(root_graph); | ||||
stream_num = context_.next_stream; | stream_num = context_.next_stream; | ||||
GELOGI("Assigned logical stream num: %ld", stream_num); | |||||
GELOGI("Assigned logical stream num: %ld.", stream_num); | |||||
return SUCCESS; | return SUCCESS; | ||||
} | } | ||||
@@ -575,7 +585,7 @@ Status LogicalStreamAllocator::DoAssign(const ComputeGraphPtr &graph, const Grap | |||||
GE_CHECK_NOTNULL(graph); | GE_CHECK_NOTNULL(graph); | ||||
NodePtr parent_node = graph->GetParentNode(); | NodePtr parent_node = graph->GetParentNode(); | ||||
if (parent_node == nullptr || parent_node->GetOpDesc() == nullptr) { | |||||
if ((parent_node == nullptr) || (parent_node->GetOpDesc() == nullptr)) { | |||||
context_.default_stream = kInvalidStream; | context_.default_stream = kInvalidStream; | ||||
} else { | } else { | ||||
context_.default_stream = parent_node->GetOpDesc()->GetStreamId(); | context_.default_stream = parent_node->GetOpDesc()->GetStreamId(); | ||||
@@ -597,10 +607,10 @@ Status LogicalStreamAllocator::DoAssign(const ComputeGraphPtr &graph, const Grap | |||||
return status; | return status; | ||||
} | } | ||||
GELOGD("Subgraphs of graph %s.", graph->GetName().c_str()); | |||||
GELOGD("Subgraphs of graph %s", graph->GetName().c_str()); | |||||
for (const auto &subgraph : subgraphs) { | for (const auto &subgraph : subgraphs) { | ||||
if (subgraph != nullptr) { | if (subgraph != nullptr) { | ||||
GELOGD("subgraph: %s.", subgraph->name.c_str()); | |||||
GELOGD("subgraph: %s", subgraph->name.c_str()); | |||||
} | } | ||||
} | } | ||||
@@ -664,9 +674,9 @@ Status LogicalStreamAllocator::RunPasses(const ComputeGraphPtr &graph, const vec | |||||
Status status = pass->Run(graph, subgraphs, context_); | Status status = pass->Run(graph, subgraphs, context_); | ||||
if (status == SUCCESS) { | if (status == SUCCESS) { | ||||
GELOGD("Stream pass %s return SUCCESS", pass->GetName().c_str()); | |||||
GELOGD("Stream pass %s return SUCCESS.", pass->GetName().c_str()); | |||||
} else if (status == NOT_CHANGED) { | } else if (status == NOT_CHANGED) { | ||||
GELOGD("Stream pass %s return NOT_CHANGED", pass->GetName().c_str()); | |||||
GELOGD("Stream pass %s return NOT_CHANGED.", pass->GetName().c_str()); | |||||
} else { | } else { | ||||
GELOGE(status, "Stream pass %s failed.", pass->GetName().c_str()); | GELOGE(status, "Stream pass %s failed.", pass->GetName().c_str()); | ||||
return status; | return status; | ||||
@@ -686,7 +696,7 @@ void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &gra | |||||
auto op_desc = node->GetOpDesc(); | auto op_desc = node->GetOpDesc(); | ||||
if (op_desc != nullptr) { | if (op_desc != nullptr) { | ||||
int64_t stream_id = op_desc->GetStreamId(); | int64_t stream_id = op_desc->GetStreamId(); | ||||
if (stream_id != kInvalidStream && stream_id < stream_num) { | |||||
if ((stream_id != kInvalidStream) && (stream_id < stream_num)) { | |||||
stream_has_node[stream_id] = true; | stream_has_node[stream_id] = true; | ||||
} | } | ||||
} | } | ||||
@@ -695,10 +705,10 @@ void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &gra | |||||
context_.next_stream = 0; | context_.next_stream = 0; | ||||
vector<int64_t> old_to_new_streams(stream_num, kInvalidStream); | vector<int64_t> old_to_new_streams(stream_num, kInvalidStream); | ||||
for (size_t old_stream = 0; old_stream < stream_has_node.size(); ++old_stream) { | |||||
for (size_t old_stream = 0; old_stream < stream_has_node.size(); old_stream++) { | |||||
if (stream_has_node[old_stream]) { | if (stream_has_node[old_stream]) { | ||||
old_to_new_streams[old_stream] = context_.next_stream; | old_to_new_streams[old_stream] = context_.next_stream; | ||||
++context_.next_stream; | |||||
context_.next_stream++; | |||||
} | } | ||||
} | } | ||||
@@ -706,7 +716,7 @@ void LogicalStreamAllocator::RefreshContinuousStreams(const ComputeGraphPtr &gra | |||||
auto op_desc = node->GetOpDesc(); | auto op_desc = node->GetOpDesc(); | ||||
if (op_desc != nullptr) { | if (op_desc != nullptr) { | ||||
int64_t stream_id = op_desc->GetStreamId(); | int64_t stream_id = op_desc->GetStreamId(); | ||||
if (stream_id != kInvalidStream && stream_id < stream_num) { | |||||
if ((stream_id != kInvalidStream) && (stream_id < stream_num)) { | |||||
op_desc->SetStreamId(old_to_new_streams[stream_id]); | op_desc->SetStreamId(old_to_new_streams[stream_id]); | ||||
} | } | ||||
} | } | ||||
@@ -40,7 +40,7 @@ static bool BlockComparator(const Block *left, const Block *right) { | |||||
} | } | ||||
bool CanMerge(Block *block) { | bool CanMerge(Block *block) { | ||||
if (block == nullptr || block->allocated || !block->IsSplit()) { | |||||
if ((block == nullptr) || block->allocated || !block->IsSplit()) { | |||||
return false; | return false; | ||||
} | } | ||||
return true; | return true; | ||||
@@ -52,7 +52,7 @@ size_t GetBinIndex(size_t size) { | |||||
if (size <= range) { | if (size <= range) { | ||||
break; | break; | ||||
} | } | ||||
++index; | |||||
index++; | |||||
} | } | ||||
if (index > kNumBins - 1) { | if (index > kNumBins - 1) { | ||||
index = kNumBins - 1; | index = kNumBins - 1; | ||||
@@ -95,17 +95,17 @@ void IncreaseCount(std::map<size_t, size_t> &count, size_t size) { | |||||
} | } | ||||
CachingAllocator::CachingAllocator(rtMemType_t memory_type) : memory_type_(memory_type), memory_allocator_(nullptr) { | CachingAllocator::CachingAllocator(rtMemType_t memory_type) : memory_type_(memory_type), memory_allocator_(nullptr) { | ||||
for (uint32_t i = 0; i < kNumBins; ++i) { | |||||
for (uint32_t i = 0; i < kNumBins; i++) { | |||||
free_block_bins_[i] = nullptr; | free_block_bins_[i] = nullptr; | ||||
} | } | ||||
} | } | ||||
Status CachingAllocator::Initialize(uint32_t device_id) { | Status CachingAllocator::Initialize(uint32_t device_id) { | ||||
GELOGI("Device id %u.", device_id); | |||||
GELOGI("Device id %u", device_id); | |||||
// when redo Initialize free old memory | // when redo Initialize free old memory | ||||
FreeBlocks(); | FreeBlocks(); | ||||
std::lock_guard<std::recursive_mutex> lock(mutex_); | std::lock_guard<std::recursive_mutex> lock(mutex_); | ||||
for (uint32_t i = 0; i < kNumBins; ++i) { | |||||
for (uint32_t i = 0; i < kNumBins; i++) { | |||||
if (free_block_bins_[i] != nullptr) { | if (free_block_bins_[i] != nullptr) { | ||||
continue; | continue; | ||||
} | } | ||||
@@ -124,14 +124,14 @@ Status CachingAllocator::Initialize(uint32_t device_id) { | |||||
} | } | ||||
void CachingAllocator::Finalize(uint32_t device_id) { | void CachingAllocator::Finalize(uint32_t device_id) { | ||||
GELOGI("Device id %u.", device_id); | |||||
GELOGI("Device id %u", device_id); | |||||
PrintStatics(); | PrintStatics(); | ||||
FreeBlocks(); | FreeBlocks(); | ||||
FreeBlockBins(); | FreeBlockBins(); | ||||
} | } | ||||
uint8_t *CachingAllocator::Malloc(size_t size, uint8_t *org_ptr, uint32_t device_id) { | uint8_t *CachingAllocator::Malloc(size_t size, uint8_t *org_ptr, uint32_t device_id) { | ||||
GELOGI("Start malloc pool memory, size = %zu, device id = %u.", size, device_id); | |||||
GELOGI("Start malloc pool memory, size = %zu, device id = %u", size, device_id); | |||||
uint8_t *ptr = nullptr; | uint8_t *ptr = nullptr; | ||||
size = GetBlockSize(size); | size = GetBlockSize(size); | ||||
Block *block = FindFreeBlock(size, org_ptr, device_id); | Block *block = FindFreeBlock(size, org_ptr, device_id); | ||||
@@ -152,7 +152,7 @@ uint8_t *CachingAllocator::Malloc(size_t size, uint8_t *org_ptr, uint32_t device | |||||
} | } | ||||
Status CachingAllocator::Free(uint8_t *ptr, uint32_t device_id) { | Status CachingAllocator::Free(uint8_t *ptr, uint32_t device_id) { | ||||
GELOGI("Free device id = %u.", device_id); | |||||
GELOGI("Free device id = %u", device_id); | |||||
if (ptr == nullptr) { | if (ptr == nullptr) { | ||||
GELOGE(PARAM_INVALID, "Invalid memory pointer"); | GELOGE(PARAM_INVALID, "Invalid memory pointer"); | ||||
return ge::PARAM_INVALID; | return ge::PARAM_INVALID; | ||||
@@ -171,10 +171,10 @@ Status CachingAllocator::Free(uint8_t *ptr, uint32_t device_id) { | |||||
} | } | ||||
void CachingAllocator::FreeBlock(Block *block) { | void CachingAllocator::FreeBlock(Block *block) { | ||||
if (block == nullptr || !block->allocated) { | |||||
if ((block == nullptr) || !block->allocated) { | |||||
return; | return; | ||||
} | } | ||||
GELOGI("Free block size = %zu.", block->size); | |||||
GELOGI("Free block size = %zu", block->size); | |||||
std::lock_guard<std::recursive_mutex> lock(mutex_); | std::lock_guard<std::recursive_mutex> lock(mutex_); | ||||
block->allocated = false; | block->allocated = false; | ||||
@@ -227,7 +227,7 @@ Block *CachingAllocator::FindFreeBlock(size_t size, uint8_t *org_ptr, uint32_t d | |||||
Block *block = *it; | Block *block = *it; | ||||
bin->erase(it); | bin->erase(it); | ||||
if (block != nullptr) { | if (block != nullptr) { | ||||
GELOGI("Find block size = %zu.", block->size); | |||||
GELOGI("Find block size = %zu", block->size); | |||||
if (ShouldSplit(block, size)) { | if (ShouldSplit(block, size)) { | ||||
block = SplitBlock(block, size, *bin, device_id); | block = SplitBlock(block, size, *bin, device_id); | ||||
} | } | ||||
@@ -235,7 +235,7 @@ Block *CachingAllocator::FindFreeBlock(size_t size, uint8_t *org_ptr, uint32_t d | |||||
if (block->ptr != nullptr) { | if (block->ptr != nullptr) { | ||||
block->allocated = true; | block->allocated = true; | ||||
allocated_blocks_[block->ptr] = block; | allocated_blocks_[block->ptr] = block; | ||||
GELOGI("Malloc device id = %u, size= %zu.", device_id, size); | |||||
GELOGI("Malloc device id = %u, size= %zu", device_id, size); | |||||
} | } | ||||
} | } | ||||
@@ -265,7 +265,7 @@ Block *CachingAllocator::SplitBlock(Block *block, size_t size, BlockBin &bin, ui | |||||
} | } | ||||
Status CachingAllocator::TryExtendCache(size_t size, uint32_t device_id) { | Status CachingAllocator::TryExtendCache(size_t size, uint32_t device_id) { | ||||
GELOGI("Try to extend cache. size = %zu, device id = %u.", size, device_id); | |||||
GELOGI("Try to extend cache. size = %zu, device id = %u", size, device_id); | |||||
auto memory_size = GetAllocationSize(size); | auto memory_size = GetAllocationSize(size); | ||||
const std::string purpose = "Memory for caching."; | const std::string purpose = "Memory for caching."; | ||||
auto memory_addr = memory_allocator_->MallocMemory(purpose, memory_size, device_id); | auto memory_addr = memory_allocator_->MallocMemory(purpose, memory_size, device_id); | ||||
@@ -302,7 +302,7 @@ Status CachingAllocator::AddToBlockBin(uint8_t *ptr, size_t size, uint32_t devic | |||||
return ge::FAILED; | return ge::FAILED; | ||||
} | } | ||||
GELOGI("Block size = %zu.", size); | |||||
GELOGI("Block size = %zu", size); | |||||
block->ptr = ptr; | block->ptr = ptr; | ||||
block->size = size; | block->size = size; | ||||
@@ -313,10 +313,10 @@ Status CachingAllocator::AddToBlockBin(uint8_t *ptr, size_t size, uint32_t devic | |||||
} | } | ||||
size_t CachingAllocator::FreeCachedBlocks() { | size_t CachingAllocator::FreeCachedBlocks() { | ||||
GELOGI("Free cached blocks."); | |||||
GELOGI("Free cached blocks"); | |||||
std::lock_guard<std::recursive_mutex> lock(mutex_); | std::lock_guard<std::recursive_mutex> lock(mutex_); | ||||
size_t free_cached_memory_size = 0; | size_t free_cached_memory_size = 0; | ||||
for (uint32_t i = 0; i < kNumBins; ++i) { | |||||
for (uint32_t i = 0; i < kNumBins; i++) { | |||||
auto pool = free_block_bins_[i]; | auto pool = free_block_bins_[i]; | ||||
if (pool == nullptr) { | if (pool == nullptr) { | ||||
continue; | continue; | ||||
@@ -324,7 +324,8 @@ size_t CachingAllocator::FreeCachedBlocks() { | |||||
for (auto it = pool->begin(); it != pool->end();) { | for (auto it = pool->begin(); it != pool->end();) { | ||||
Block *block = *it; | Block *block = *it; | ||||
// free block memory that has not been split | // free block memory that has not been split | ||||
if ((block != nullptr) && (block->ptr != nullptr) && (block->prev == nullptr) && (block->next == nullptr) && | |||||
if ((block != nullptr) && (block->ptr != nullptr) && | |||||
(block->prev == nullptr) && (block->next == nullptr) && | |||||
(memory_allocator_->FreeMemory(block->ptr) == ge::SUCCESS)) { | (memory_allocator_->FreeMemory(block->ptr) == ge::SUCCESS)) { | ||||
auto itcount = malloced_memory_.find(block->size); | auto itcount = malloced_memory_.find(block->size); | ||||
free_cached_memory_size += block->size; | free_cached_memory_size += block->size; | ||||
@@ -345,7 +346,7 @@ size_t CachingAllocator::FreeCachedBlocks() { | |||||
} | } | ||||
void CachingAllocator::FreeBlocks() { | void CachingAllocator::FreeBlocks() { | ||||
GELOGI("Free blocks"); | |||||
GELOGI("Free blocks."); | |||||
std::lock_guard<std::recursive_mutex> lock(mutex_); | std::lock_guard<std::recursive_mutex> lock(mutex_); | ||||
// free allocated blocks and put to cache | // free allocated blocks and put to cache | ||||
for (auto &it : allocated_blocks_) { | for (auto &it : allocated_blocks_) { | ||||
@@ -356,9 +357,9 @@ void CachingAllocator::FreeBlocks() { | |||||
} | } | ||||
void CachingAllocator::FreeBlockBins() { | void CachingAllocator::FreeBlockBins() { | ||||
GELOGI("Free block bins"); | |||||
GELOGI("Free block bins."); | |||||
std::lock_guard<std::recursive_mutex> lock(mutex_); | std::lock_guard<std::recursive_mutex> lock(mutex_); | ||||
for (uint32_t i = 0; i < kNumBins; ++i) { | |||||
for (uint32_t i = 0; i < kNumBins; i++) { | |||||
if (free_block_bins_[i] != nullptr) { | if (free_block_bins_[i] != nullptr) { | ||||
delete free_block_bins_[i]; | delete free_block_bins_[i]; | ||||
free_block_bins_[i] = nullptr; | free_block_bins_[i] = nullptr; | ||||
@@ -367,9 +368,9 @@ void CachingAllocator::FreeBlockBins() { | |||||
} | } | ||||
void PrintCount(std::map<size_t, size_t> &count, const std::string &name, size_t total_size, size_t total_count) { | void PrintCount(std::map<size_t, size_t> &count, const std::string &name, size_t total_size, size_t total_count) { | ||||
GELOGI("%6s total[size:%10zu count:%10zu]", name.c_str(), total_size, total_count); | |||||
GELOGI("%6s total[size:%10zu count:%10zu].", name.c_str(), total_size, total_count); | |||||
for (auto &it : count) { | for (auto &it : count) { | ||||
GELOGI(" |- block[size:%10zu count:%10zu]", it.first, it.second); | |||||
GELOGI(" |- block[size:%10zu count:%10zu].", it.first, it.second); | |||||
} | } | ||||
} | } | ||||
@@ -383,20 +384,20 @@ void CachingAllocator::PrintStatics() { | |||||
size_t total_free_count = 0; | size_t total_free_count = 0; | ||||
size_t total_malloc_size = 0; | size_t total_malloc_size = 0; | ||||
size_t total_malloc_count = 0; | size_t total_malloc_count = 0; | ||||
std::map<size_t, size_t> using_block; | |||||
std::map<size_t, size_t> free_block; | |||||
std::map<size_t, size_t> malloc_block; | |||||
std::map<size_t, size_t> using_block_stat; | |||||
std::map<size_t, size_t> free_block_stat; | |||||
std::map<size_t, size_t> malloc_block_stat; | |||||
do { | do { | ||||
std::lock_guard<std::recursive_mutex> lock(mutex_); | std::lock_guard<std::recursive_mutex> lock(mutex_); | ||||
for (uint32_t i = 0; i < kNumBins; ++i) { | |||||
for (uint32_t i = 0; i < kNumBins; i++) { | |||||
auto pool = free_block_bins_[i]; | auto pool = free_block_bins_[i]; | ||||
if (pool == nullptr) { | if (pool == nullptr) { | ||||
continue; | continue; | ||||
} | } | ||||
for (auto it = pool->begin(); it != pool->end(); ++it) { | |||||
for (auto it = pool->begin(); it != pool->end(); it++) { | |||||
if ((*it) != nullptr) { | if ((*it) != nullptr) { | ||||
total_free_size += (*it)->size; | total_free_size += (*it)->size; | ||||
IncreaseCount(free_block, (*it)->size); | |||||
IncreaseCount(free_block_stat, (*it)->size); | |||||
total_free_count++; | total_free_count++; | ||||
} | } | ||||
} | } | ||||
@@ -405,7 +406,7 @@ void CachingAllocator::PrintStatics() { | |||||
for (auto &it : allocated_blocks_) { | for (auto &it : allocated_blocks_) { | ||||
if (it.second != nullptr) { | if (it.second != nullptr) { | ||||
total_using_size += it.second->size; | total_using_size += it.second->size; | ||||
IncreaseCount(using_block, it.second->size); | |||||
IncreaseCount(using_block_stat, it.second->size); | |||||
total_using_count++; | total_using_count++; | ||||
} | } | ||||
} | } | ||||
@@ -413,12 +414,12 @@ void CachingAllocator::PrintStatics() { | |||||
for (auto &it : malloced_memory_) { | for (auto &it : malloced_memory_) { | ||||
total_malloc_size += it.first * it.second; | total_malloc_size += it.first * it.second; | ||||
total_malloc_count += it.second; | total_malloc_count += it.second; | ||||
malloc_block[it.first] = it.second; | |||||
malloc_block_stat[it.first] = it.second; | |||||
} | } | ||||
} while (0); | } while (0); | ||||
PrintCount(malloc_block, "Malloc", total_malloc_size, total_malloc_count); | |||||
PrintCount(using_block, "Using", total_using_size, total_using_count); | |||||
PrintCount(free_block, "Free", total_free_size, total_free_count); | |||||
PrintCount(malloc_block_stat, "Malloc", total_malloc_size, total_malloc_count); | |||||
PrintCount(using_block_stat, "Using", total_using_size, total_using_count); | |||||
PrintCount(free_block_stat, "Free", total_free_size, total_free_count); | |||||
} | } | ||||
} // namespace ge | } // namespace ge |