IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

enabled exec_graph and pc in deserialized model (#15975)

Browse Source
This commit is contained in:
Eddy Kim 2023-02-28 03:14:04 +09:00 committed by GitHub
parent 2ced2ad929
commit d2a5be0ab8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 105 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/plugins/intel_gpu/include/intel_gpu/graph/network.hpp
View File

@ -259,6 +259,8 @@ private:
std::list<std::shared_ptr<primitive_inst>> _data_outputs;
variables_states_map _variables_states;
std::vector<std::shared_ptr<primitive_inst>> _variable_state_primitives;
program::primitives_info _prims_info;
std::map<primitive_id, primitive_id> _ext_id_mapping;
std::unordered_map<primitive_id, event::ptr> _events;
output_chains_map _output_chains;

58
src/plugins/intel_gpu/src/graph/network.cpp
View File

@ -461,6 +461,40 @@ network::network(cldnn::BinaryInputBuffer& ib, const ExecutionConfig& config, st
}
add_default_output_chains();
size_t prims_info_size;
ib >> prims_info_size;
for (size_t i = 0; i < prims_info_size; i++) {
primitive_id original_id;
std::string type_id;
primitive::primitive_id_arr c_dependencies;
primitive::primitive_id_arr c_users;
primitive::primitive_id_arr c_fused_ids;
layout output_layout;
std::string layout_str;
std::string kernel_id;
data_types runtime_precision;
bool is_cpu;
int exec_id;
ib >> original_id;
ib >> type_id;
ib >> c_dependencies;
ib >> c_users;
ib >> c_fused_ids;
ib >> output_layout;
ib >> layout_str;
ib >> kernel_id;
ib >> make_data(&runtime_precision, sizeof(data_types));
ib >> is_cpu;
ib >> exec_id;
primitive_info prim_info(original_id, type_id, c_dependencies, c_users, c_fused_ids,
output_layout, layout_str, kernel_id, runtime_precision, is_cpu, exec_id);
_prims_info.emplace_back(prim_info);
}
ib >> _ext_id_mapping;
}
network::~network() {
@ -544,6 +578,24 @@ void network::save(cldnn::BinaryOutputBuffer& ob) {
for (const auto& p_inst : _variable_state_primitives) {
ob << p_inst->id();
}
auto& prims_info = get_primitives_info();
ob << prims_info.size();
for (auto& prim_info : prims_info) {
ob << prim_info.original_id;
ob << prim_info.type_id;
ob << prim_info.c_dependencies;
ob << prim_info.c_users;
ob << prim_info.c_fused_ids;
ob << prim_info.output_layout;
ob << prim_info.layout_str;
ob << prim_info.kernel_id;
ob << make_data(&prim_info.runtime_precision, sizeof(data_types));
ob << prim_info.is_cpu;
ob << prim_info.exec_id;
}
ob << get_ext_id_mapping();
}
network::ptr network::allocate_network(stream::ptr stream, program::ptr program, bool is_internal, bool is_primary_stream) {
@ -1123,7 +1175,7 @@ std::vector<primitive_id> network::get_all_primitive_org_ids() const {
}
const program::primitives_info& network::get_primitives_info() const {
return _program->get_primitives_info();
return (_program == nullptr) ? _prims_info : _program->get_primitives_info();
}
const program::graph_optimizer_info& network::get_optimizer_passes_info() const {
@ -1131,6 +1183,10 @@ const program::graph_optimizer_info& network::get_optimizer_passes_info() const
}
std::map<primitive_id, primitive_id> network::get_ext_id_mapping() const {
if (_program == nullptr) {
return _ext_id_mapping;
}
std::map<primitive_id, primitive_id> result;
for (auto& prim : _primitives) {
result.emplace(prim.first, prim.second->get_node().get_primitive()->origin_op_name);

47
src/plugins/intel_gpu/src/plugin/graph.cpp
View File

@ -82,6 +82,23 @@ Graph::Graph(cldnn::BinaryInputBuffer &ib, RemoteContextImpl::Ptr context, const
m_program->AddVariableStateInfo(variablesStateInfo.first, *variablesStateInfo.second.begin());
}
ib >> primitiveIDs;
ib >> prevPrimitiveIDs;
ib >> profilingIDs;
{
size_t perfMap_size;
ib >> perfMap_size;
for (size_t i = 0; i < perfMap_size; ++i) {
cldnn::primitive_id prim_id;
ib >> prim_id;
perfMap[prim_id].first = prim_id;
auto& perfEntry = perfMap[prim_id].second;
ib >> perfEntry.layerType;
ib >> cldnn::make_data(&perfEntry.status, sizeof(InferenceEngine::InferenceEngineProfileInfo::LayerStatus));
perfEntry.cpu_uSec = perfEntry.realTime_uSec = 0;
ib >> perfEntry.isCPU;
ib >> perfEntry.parentPrimitive;
}
}
ib >> outputDims;
size_t num_networks;
@ -502,6 +519,18 @@ void Graph::Export(cldnn::BinaryOutputBuffer &ob) {
ob << m_program->inputLayouts;
ob << m_program->GetVariablesStatesInfo();
ob << primitiveIDs;
ob << prevPrimitiveIDs;
ob << profilingIDs;
{
ob << perfMap.size();
for (auto& perf_item : perfMap) {
ob << perf_item.first;
ob << perf_item.second.second.layerType;
ob << cldnn::make_data(&perf_item.second.second.status, sizeof(InferenceEngine::InferenceEngineProfileInfo::LayerStatus));
ob << perf_item.second.second.isCPU;
ob << perf_item.second.second.parentPrimitive;
}
}
ob << outputDims;
ob << m_networks.size();
@ -597,6 +626,13 @@ std::map<std::string, InferenceEngine::InferenceEngineProfileInfo> Graph::GetPer
auto executedPrimitives = GetNetwork()->get_executed_primitives();
auto primitivesInfo = GetNetwork()->get_primitives_info();
auto extIdMap = GetNetwork()->get_ext_id_mapping();
std::map<std::string, std::string> implementation_info;
if (GetNetwork()->get_program() == nullptr) {
for (auto& pi : primitivesInfo) {
implementation_info[pi.original_id] = pi.kernel_id;
}
}
auto getUpperCaseName = [](std::string name) {
std::vector<char> res;
@ -641,7 +677,16 @@ std::map<std::string, InferenceEngine::InferenceEngineProfileInfo> Graph::GetPer
static const std::string cpuExecType("CPU");
cpuExecType.copy(extPerfEntry.exec_type, cpuExecType.length()); // Override execType as CPU
} else {
std::string impl = GetNetwork()->get_implementation_info(primId);
std::string impl;
if (GetNetwork()->get_program() != nullptr) {
impl = GetNetwork()->get_implementation_info(primId);
} else {
if (implementation_info.find(primId) != implementation_info.end()) {
impl = implementation_info[primId];
} else {
impl = "undef";
}
}
impl.copy(extPerfEntry.exec_type, impl.length());
}