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

[GPU] updates to use a graph local net_id for model caching (#18372)

Browse Source 
* newly added local_net_id

* updated calls to network ctors

* updated to use uint32_t for local_net_id

* added comments for _local_net_id
This commit is contained in:
Eddy Kim
2023-07-06 18:34:08 -07:00
committed by GitHub
parent 24bfb29eb8
commit 58d1fc3c6b
10 changed files with 23 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/plugins/intel_gpu/include/intel_gpu/graph/network.hpp
View File

@@ -92,8 +92,8 @@ public:
network(program::ptr program, stream::ptr stream, uint16_t stream_id);
network(cldnn::BinaryInputBuffer& ifs, stream::ptr stream, engine& engine, bool is_primary_stream = true);
network(cldnn::BinaryInputBuffer& ifs, const ExecutionConfig& config, stream::ptr stream, engine& engine, bool is_primary_stream = true);
network(cldnn::BinaryInputBuffer& ifs, stream::ptr stream, engine& engine, bool is_primary_stream, uint32_t local_net_id);
network(cldnn::BinaryInputBuffer& ifs, const ExecutionConfig& config, stream::ptr stream, engine& engine, bool is_primary_stream, uint32_t local_net_id);
~network();
@@ -215,6 +215,7 @@ public:
void configure_primitives_second_output();
void build_insts_deps();
uint32_t get_id() const { return net_id; }
uint32_t get_local_id() const { return _local_net_id; }
stream& get_stream() const { return *_stream; }
stream::ptr get_stream_ptr() const { return _stream; }
bool is_internal() const { return _internal; }
@@ -254,6 +255,8 @@ private:
bool _is_dynamic = false;
bool _enable_profiling = false;
bool _reset_arguments;
uint32_t _local_net_id = 0; // This is for thread-safe deserialization. 'net_id' is globally unique,
// but '_local_net_id' is unique only in each intel_gpu::Graph.
std::unordered_map<primitive_id, std::shared_ptr<primitive_inst>> _primitives;
std::vector<shared_mem_type> _in_out_shared_mem_types;

10
src/plugins/intel_gpu/include/intel_gpu/graph/serialization/binary_buffer.hpp
View File

@@ -37,7 +37,7 @@ private:
class BinaryInputBuffer : public InputBuffer<BinaryInputBuffer> {
public:
BinaryInputBuffer(std::istream& stream, engine& engine)
: InputBuffer(this, engine), _stream(stream), _impl_params(nullptr), _num_networks(0), _stream_id(0) {}
: InputBuffer(this, engine), _stream(stream), _impl_params(nullptr) {}
void read(void* const data, std::streamsize size) {
auto const read_size = _stream.rdbuf()->sgetn(reinterpret_cast<char*>(data), size);
@@ -62,18 +62,10 @@ public:
std::streampos tellg() { return _stream.tellg(); }
void seekg(std::streampos pos) { _stream.seekg(pos); }
void new_network_added() { _num_networks += 1; }
int get_num_networks() const { return _num_networks; }
void set_stream_id(uint16_t stream_id) { _stream_id = stream_id; }
uint16_t get_stream_id() const { return _stream_id; }
private:
std::istream& _stream;
void* _impl_params;
std::vector<std::unordered_map<std::string, std::shared_ptr<memory>>> _const_data_map;
int _num_networks;
uint16_t _stream_id;
};
template <typename T>

4
src/plugins/intel_gpu/src/graph/data.cpp
View File

@@ -87,7 +87,7 @@ void data_inst::load(BinaryInputBuffer& ib) {
ib >> make_data(&data_size, sizeof(size_t));
if (!get_network().is_primary_stream()) {
_outputs[0] = ib.getConstData(get_network_id() - (ib.get_num_networks() * ib.get_stream_id()), id());
_outputs[0] = ib.getConstData(_network.get_local_id(), id());
auto pos = ib.tellg();
pos += data_size;
ib.seekg(pos);
@@ -103,7 +103,7 @@ void data_inst::load(BinaryInputBuffer& ib) {
_outputs[0]->copy_from(get_network().get_stream(), _buf.data());
}
ib.addConstData(get_network_id(), id(), _outputs[0]);
ib.addConstData(_network.get_local_id(), id(), _outputs[0]);
}
}

2
src/plugins/intel_gpu/src/graph/loop.cpp
View File

@@ -529,7 +529,7 @@ void loop_inst::load(BinaryInputBuffer& ib) {
ib >> _condition_id;
ib >> _num_iteration_id;
ib >> _max_iteration;
body_network = std::make_shared<cldnn::network>(ib, get_network().get_stream_ptr(), get_network().get_engine(), get_network().is_primary_stream());
body_network = std::make_shared<cldnn::network>(ib, get_network().get_stream_ptr(), get_network().get_engine(), get_network().is_primary_stream(), 0);
}
} // namespace cldnn

11
src/plugins/intel_gpu/src/graph/network.cpp
View File

@@ -359,10 +359,10 @@ network::network(program::ptr program, uint16_t stream_id)
network::network(program::ptr program, stream::ptr stream, uint16_t stream_id)
: network(program, program->get_config(), stream, false, stream_id == 0) {}
network::network(cldnn::BinaryInputBuffer& ib, stream::ptr stream, engine& engine, bool is_primary_stream)
: network(ib, ExecutionConfig{}, stream, engine, is_primary_stream) {}
network::network(cldnn::BinaryInputBuffer& ib, stream::ptr stream, engine& engine, bool is_primary_stream, uint32_t local_net_id)
: network(ib, ExecutionConfig{}, stream, engine, is_primary_stream, local_net_id) {}
network::network(cldnn::BinaryInputBuffer& ib, const ExecutionConfig& config, stream::ptr stream, engine& engine, bool is_primary_stream)
network::network(cldnn::BinaryInputBuffer& ib, const ExecutionConfig& config, stream::ptr stream, engine& engine, bool is_primary_stream, uint32_t local_net_id)
: _program(nullptr)
, _config(config)
, _engine(engine)
@@ -370,10 +370,9 @@ network::network(cldnn::BinaryInputBuffer& ib, const ExecutionConfig& config, st
, _memory_pool(new memory_pool(engine))
, _internal(false)
, _is_primary_stream(is_primary_stream)
, _reset_arguments(true) {
, _reset_arguments(true)
, _local_net_id(local_net_id) {
net_id = get_unique_net_id();
if (is_primary_stream)
ib.new_network_added();
kernels_cache kernels_cache(get_engine(), config, 0, nullptr, {""});
ib >> kernels_cache;

5
src/plugins/intel_gpu/src/plugin/graph.cpp
View File

@@ -106,9 +106,8 @@ Graph::Graph(cldnn::BinaryInputBuffer &ib, RemoteContextImpl::Ptr context, const
size_t num_networks;
ib >> num_networks;
for (size_t i = 0; i < num_networks; ++i) {
ib.set_stream_id(m_stream_id);
m_networks.emplace_back(std::make_shared<cldnn::network>(ib, get_engine().create_stream(config), get_engine(), m_stream_id == 0));
for (uint32_t i = 0; i < num_networks; ++i) {
m_networks.emplace_back(std::make_shared<cldnn::network>(ib, get_engine().create_stream(config), get_engine(), m_stream_id == 0, i));
}
}

2
src/plugins/intel_gpu/tests/unit/test_cases/gemm_gpu_test.cpp
View File

@@ -138,7 +138,7 @@ public:
{
std::istream in_mem(&mem_buf);
BinaryInputBuffer ib = BinaryInputBuffer(in_mem, engine);
network = std::make_shared<cldnn::network>(ib, get_test_stream_ptr(), engine);
network = std::make_shared<cldnn::network>(ib, get_test_stream_ptr(), engine, true, 0);
}
} else {
network = std::make_shared<cldnn::network>(engine, tp, get_test_default_config(engine));

2
src/plugins/intel_gpu/tests/unit/test_cases/proposal_cpu_test.cpp
View File

@@ -101,7 +101,7 @@ TestRunnerProposal<Dtype, ImInfoType>::TestRunnerProposal(cldnn::tensor image_in
{
std::istream in_mem(&mem_buf);
BinaryInputBuffer ib = BinaryInputBuffer(in_mem, get_test_engine());
_network.reset(new network(ib, get_test_stream_ptr(), get_test_engine()));
_network.reset(new network(ib, get_test_stream_ptr(), get_test_engine(), true, 0));
}
} else {
_network.reset(new network(get_test_engine(), _topology));

10
src/plugins/intel_gpu/tests/unit/test_cases/streams_test.cpp
View File

@@ -93,10 +93,9 @@ public:
std::istream in_mem(&mem_buf);
BinaryInputBuffer ib = BinaryInputBuffer(in_mem, engine);
auto pos = ib.tellg();
network0 = std::make_shared<cldnn::network>(ib, get_test_stream_ptr(), engine, true);
network0 = std::make_shared<cldnn::network>(ib, get_test_stream_ptr(), engine, true, 0);
ib.seekg(pos);
ib.set_stream_id(1);
network1 = std::make_shared<cldnn::network>(ib, get_test_stream_ptr(), engine, false);
network1 = std::make_shared<cldnn::network>(ib, get_test_stream_ptr(), engine, false, 0);
}
}
} else {
@@ -193,13 +192,12 @@ public:
{
std::istream in_mem0(&mem_buf0);
BinaryInputBuffer ib0 = BinaryInputBuffer(in_mem0, engine);
network0 = std::make_shared<cldnn::network>(ib0, get_test_stream_ptr(), engine, false);
network0 = std::make_shared<cldnn::network>(ib0, get_test_stream_ptr(), engine, false, 0);
}
{
std::istream in_mem1(&mem_buf1);
BinaryInputBuffer ib1 = BinaryInputBuffer(in_mem1, engine);
ib1.set_stream_id(1);
network1 = std::make_shared<cldnn::network>(ib1, get_test_stream_ptr(), engine, true);
network1 = std::make_shared<cldnn::network>(ib1, get_test_stream_ptr(), engine, true, 0);
}
}
} else {

2
src/plugins/intel_gpu/tests/unit/test_utils/test_utils.h
View File

@@ -771,7 +771,7 @@ inline cldnn::network::ptr get_network(cldnn::engine& engine,
{
std::istream in_mem(&mem_buf);
cldnn::BinaryInputBuffer ib = cldnn::BinaryInputBuffer(in_mem, engine);
network = std::make_shared<cldnn::network>(ib, config, stream, engine);
network = std::make_shared<cldnn::network>(ib, config, stream, engine, true, 0);
}
} else {
network = std::make_shared<cldnn::network>(engine, topology, config);