Skip to content

N
ngraph
Commit 013c2381 authored by Sergey Shalnov's avatar Sergey Shalnov Committed by Robert Kimball
Browse files
Options

IntelGPU backend: Separate backend and executable classes (#2447)

parent 65141c5f
Show whitespace changes
Inline Side-by-side
Showing with 428 additions and 364 deletions
src/ngraph/runtime/intelgpu/CMakeLists.txt
View file @ 013c2381
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
set(SRC set(SRC
intelgpu_backend.cpp intelgpu_backend.cpp
intelgpu_executable.cpp
intelgpu_tensor_view.cpp intelgpu_tensor_view.cpp
intelgpu_layout.cpp intelgpu_layout.cpp
intelgpu_op_batchnorm.cpp intelgpu_op_batchnorm.cpp
......
src/ngraph/runtime/intelgpu/intelgpu_backend.cpp
View file @ 013c2381
...@@ -15,8 +15,6 @@ ...@@ -15,8 +15,6 @@
//***************************************************************************** //*****************************************************************************
#include <iomanip> #include <iomanip>
#include <sys/resource.h>
#include <sys/time.h>
#include <CPP/activation.hpp> #include <CPP/activation.hpp>
#include <CPP/activation_grad.hpp> #include <CPP/activation_grad.hpp>
...@@ -37,9 +35,7 @@ ...@@ -37,9 +35,7 @@
#include <CPP/mutable_data.hpp> #include <CPP/mutable_data.hpp>
#include <CPP/permute.hpp> #include <CPP/permute.hpp>
#include <CPP/pooling.hpp> #include <CPP/pooling.hpp>
#include <CPP/reorder.hpp>
#include <CPP/reshape.hpp> #include <CPP/reshape.hpp>
#include <CPP/scale.hpp>
#include <CPP/select.hpp> #include <CPP/select.hpp>
#include <CPP/softmax.hpp> #include <CPP/softmax.hpp>
#include <CPP/topology.hpp> #include <CPP/topology.hpp>
...@@ -51,6 +47,7 @@ ...@@ -51,6 +47,7 @@
#include "ngraph/pass/nop_elimination.hpp" #include "ngraph/pass/nop_elimination.hpp"
#include "ngraph/pass/reshape_elimination.hpp" #include "ngraph/pass/reshape_elimination.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_backend.hpp" #include "ngraph/runtime/intelgpu/intelgpu_backend.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_executable.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_layout.hpp" #include "ngraph/runtime/intelgpu/intelgpu_layout.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_op_batchnorm.hpp" #include "ngraph/runtime/intelgpu/intelgpu_op_batchnorm.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_op_broadcast.hpp" #include "ngraph/runtime/intelgpu/intelgpu_op_broadcast.hpp"
...@@ -141,21 +138,6 @@ static void arguments_check(const shared_ptr<Node>& op, size_t input, size_t out ...@@ -141,21 +138,6 @@ static void arguments_check(const shared_ptr<Node>& op, size_t input, size_t out
} }
} }
static void
memory_size_check(size_t memory_size, const shared_ptr<Node>& node, const string& function_name)
{
const size_t tensor_size = shape_size(node->get_shape()) * node->get_element_type().size();
if (memory_size != tensor_size)
{
ostringstream os;
os << "IntelGPU backend failed memory check. In \"" << function_name << "\" with Node \""
<< node->get_name() << "\" and " << node->get_shape() << " mismatched memory sizes "
<< tensor_size << " and " << memory_size;
throw invalid_argument(os.str());
}
}
static const string& get_input_name(const shared_ptr<Node>& op, size_t num = 0) static const string& get_input_name(const shared_ptr<Node>& op, size_t num = 0)
{ {
return op->get_inputs().at(num).get_tensor().get_name(); return op->get_inputs().at(num).get_tensor().get_name();
...@@ -312,22 +294,6 @@ extern "C" void delete_backend(runtime::Backend* backend) ...@@ -312,22 +294,6 @@ extern "C" void delete_backend(runtime::Backend* backend)
delete backend; delete backend;
} }
static size_t get_max_memory_rss()
{
size_t result = 0;
struct rusage usage;
if (getrusage(RUSAGE_SELF, &usage) == 0)
{
result = usage.ru_maxrss; // the value is in kilobytes
// aligne result to return bytes
result *= 1000;
}
return result;
}
runtime::intelgpu::IntelGPUBackend::IntelGPUBackend() runtime::intelgpu::IntelGPUBackend::IntelGPUBackend()
{ {
bool profiling = false; bool profiling = false;
...@@ -433,7 +399,7 @@ shared_ptr<runtime::Executable> ...@@ -433,7 +399,7 @@ shared_ptr<runtime::Executable>
if (m_profile_enable) if (m_profile_enable)
{ {
consumed_memory = get_max_memory_rss(); consumed_memory = runtime::intelgpu::get_max_memory_rss();
timer_compile.start(); timer_compile.start();
} }
...@@ -1831,7 +1797,7 @@ shared_ptr<runtime::Executable> ...@@ -1831,7 +1797,7 @@ shared_ptr<runtime::Executable>
{ {
timer_compile.stop(); timer_compile.stop();
compilation_time = timer_compile.get_milliseconds(); compilation_time = timer_compile.get_milliseconds();
consumed_memory = get_max_memory_rss() - consumed_memory; consumed_memory = runtime::intelgpu::get_max_memory_rss() - consumed_memory;
} }
rc = make_shared<IntelGPUExecutable>(func, rc = make_shared<IntelGPUExecutable>(func,
...@@ -1849,102 +1815,6 @@ shared_ptr<runtime::Executable> ...@@ -1849,102 +1815,6 @@ shared_ptr<runtime::Executable>
return rc; return rc;
} }
runtime::intelgpu::IntelGPUExecutable::IntelGPUExecutable(shared_ptr<Function> func,
shared_ptr<cldnn::network> network,
bool enable_timing,
bool enable_profile,
double compilation_time,
double consumed_memory,
size_t profile_lines_limit_count)
{
m_function = func;
m_cldnn_network = network;
m_performance_counters_enabled = enable_timing;
m_profile_enable = enable_profile;
m_compilation_time = compilation_time;
m_consumed_memory = consumed_memory;
m_profile_lines_limit_count = profile_lines_limit_count;
set_parameters_and_results(*func);
}
bool runtime::intelgpu::IntelGPUExecutable::call(const vector<shared_ptr<runtime::Tensor>>& outputs,
const vector<shared_ptr<runtime::Tensor>>& inputs)
{
double mem_call_consumed = 0.0f;
stopwatch timer_call;
if (m_cldnn_network == nullptr)
{
throw runtime_error("compile() must be called before call().");
}
if (m_profile_enable)
{
mem_call_consumed = get_max_memory_rss();
timer_call.start();
}
// Process input parameters. Correctness of parameters was validated by validate_call.
// Since we have no correlation between Function::m_parameters and inputs, there is
// we try to match them by index number in vectors.
for (size_t i = 0; i < inputs.size(); i++)
{
shared_ptr<runtime::intelgpu::IntelGPUTensorView> tv =
static_pointer_cast<runtime::intelgpu::IntelGPUTensorView>(inputs[i]);
const ParameterVector& input_params = get_parameters();
const string& tensor_name = input_params[i]->get_output_tensor().get_name();
m_cldnn_network->set_input_data(tensor_name, *tv->get_data_ptr());
}
// Execute network
map<cldnn::primitive_id, cldnn::network_output> result = m_cldnn_network->execute();
// Process output parameters. Correctness of parameters was validated by validate_call.
// Since we have no correlation between Function::m_results and outputs, there is
// we try to match them by index number in vectors.
for (size_t i = 0; i < m_function->get_output_size(); i++)
{
const shared_ptr<Node>& dst_node = m_function->get_output_op(i);
const size_t dst_shape_size = shape_size(dst_node->get_shape());
// We should not touch destination memory if it is not existed
if (!dst_shape_size)
{
continue;
}
shared_ptr<runtime::intelgpu::IntelGPUTensorView> ngraph_res =
static_pointer_cast<runtime::intelgpu::IntelGPUTensorView>(outputs[i]);
const string& tensor_name = get_input_name(dst_node);
auto result_memory = result.at(tensor_name).get_memory().pointer<char>();
memory_size_check(result_memory.size(), dst_node, m_function->get_name());
ngraph_res->write(result_memory.data(), 0, result_memory.size());
}
if (m_profile_enable)
{
timer_call.stop();
mem_call_consumed = get_max_memory_rss() - mem_call_consumed;
print_call_performance(m_cldnn_network,
m_function,
m_compilation_time,
timer_call.get_milliseconds(),
m_consumed_memory,
mem_call_consumed,
get_max_memory_rss());
// Output compile time only once
m_compilation_time = 0.0;
m_consumed_memory = 0.0;
}
return true;
}
void runtime::intelgpu::IntelGPUBackend::remove_compiled_function(shared_ptr<Executable> exec) void runtime::intelgpu::IntelGPUBackend::remove_compiled_function(shared_ptr<Executable> exec)
{ {
for (auto it = cldnn_networks.begin(); it != cldnn_networks.end(); ++it) for (auto it = cldnn_networks.begin(); it != cldnn_networks.end(); ++it)
...@@ -1957,199 +1827,6 @@ void runtime::intelgpu::IntelGPUBackend::remove_compiled_function(shared_ptr<Exe ...@@ -1957,199 +1827,6 @@ void runtime::intelgpu::IntelGPUBackend::remove_compiled_function(shared_ptr<Exe
} }
} }
// The cldnn::network contains something like "generic_layer_0_Parameter_254_0" names
// This function should return "Parameter_254" from the example above
static string convert_cldnn_names(shared_ptr<Function> func, const string& cldnn_name)
{
const string key("_");
string result;
const size_t last_key = cldnn_name.rfind(key);
const size_t pre_last_key = cldnn_name.rfind(key, last_key - 1);
const size_t pre_pre_last_key = cldnn_name.rfind(key, pre_last_key - 1);
if (pre_pre_last_key == std::string::npos)
{
result = cldnn_name.substr(0, last_key);
}
else
{
result = cldnn_name.substr(pre_pre_last_key + 1, last_key - pre_pre_last_key - 1);
}
return result;
}
vector<runtime::PerformanceCounter>
runtime::intelgpu::IntelGPUExecutable::get_performance_data() const
{
vector<runtime::PerformanceCounter> rc;
if (m_cldnn_network != nullptr && m_performance_counters_enabled)
{
const map<cldnn::primitive_id, cldnn::event>& primitives =
m_cldnn_network->get_executed_primitives();
for (const auto& p : primitives)
{
// Let's generate the primitive name that matches to the name in Function
const string primitive_name = convert_cldnn_names(m_function, p.first);
size_t usec = 0;
for (const auto& q : p.second.get_profiling_info())
{
if (q.name == string("executing"))
{
usec += chrono::duration_cast<
chrono::duration<size_t, chrono::milliseconds::period>>(
q.value->value())
.count();
}
}
const runtime::PerformanceCounter perf_counter(primitive_name.c_str(), usec, 1);
rc.push_back(perf_counter);
}
}
return rc;
}
static Node* get_node_by_name(const shared_ptr<Function> func, const string& name)
{
for (shared_ptr<Node> node : func->get_ops())
{
if (node->get_name() == name)
{
return node.get();
}
}
return nullptr;
}
void runtime::intelgpu::IntelGPUExecutable::print_call_performance(
const shared_ptr<cldnn::network> network,
const shared_ptr<Function> func,
double time_compile,
double time_call,
double mem_compilation_consumed,
double mem_call_consumed,
double mem_current) const
{
struct data_item
{
string item_name;
map<string, double> item_times;
};
const string& func_name = func->get_name();
const map<cldnn::primitive_id, cldnn::event>& primitives = network->get_executed_primitives();
size_t limit_count = m_profile_lines_limit_count;
multimap<double, data_item> data;
map<string, double> total_interval_times;
double total_executing_time = 0;
size_t total_items_count = 0;
size_t max_item_name_size = 0;
ios_base::fmtflags saved_stream_flags(cout.flags()); // Save stream flags to restore them later
if (m_profile_lines_limit_count > 0)
{
// Extract profiling statistic, calculate summary and sort
for (auto& prim : primitives)
{
double executing_time = 0;
data_item item;
item.item_name = prim.first;
max_item_name_size = max(max_item_name_size, prim.first.size());
for (auto& prof_info : prim.second.get_profiling_info())
{
const string& interval_name = prof_info.name;
double interval =
chrono::duration_cast<chrono::duration<double, chrono::milliseconds::period>>(
prof_info.value->value())
.count();
item.item_times[interval_name] = interval;
// Get the Key time to sort by
if (interval_name == "executing")
{
executing_time += interval;
}
// Accumulate total time for each interval
if (total_interval_times.find(interval_name) == total_interval_times.end())
{
total_interval_times[interval_name] = interval;
}
else
{
total_interval_times[interval_name] += interval;
}
}
data.emplace(executing_time, item);
total_executing_time += executing_time;
++total_items_count;
}
// Print statistic for each primitive in the cldnn::network
for (auto it = data.rbegin(); (it != data.rend()) && (limit_count > 0); ++it, --limit_count)
{
const string ngraph_node_name = convert_cldnn_names(func, it->second.item_name);
const Node* ngraph_node = get_node_by_name(func, ngraph_node_name);
cout << func_name << delim << setw(max_item_name_size) << it->second.item_name << delim
<< "time(ms)" << delim << scientific << setprecision(2) << it->first;
for (auto item : it->second.item_times)
{
cout << delim << item.first << "(ms)" << delim << item.second;
}
cout << delim << ngraph_node_name;
if (ngraph_node) // it might be initialized by nullptr
{
// print all input shapes for the Node
size_t arg_idx = 0;
for (const descriptor::Input& op_input : ngraph_node->get_inputs())
{
cout << delim << op_input.get_element_type().c_type_string() << " input"
<< arg_idx << vector_to_string(op_input.get_shape());
++arg_idx;
}
// print all output shapes for the Node
arg_idx = 0;
for (const descriptor::Output& op_output : ngraph_node->get_outputs())
{
cout << delim << op_output.get_element_type().c_type_string() << " output"
<< arg_idx << vector_to_string(op_output.get_shape());
++arg_idx;
}
}
cout << "\n";
}
// Print bottom line summary
const string total_items_count_string = "Total(cldnn " + to_string(total_items_count) +
", ngraph " + to_string(func->get_ops().size()) +
")";
cout << func_name << delim << setw(max_item_name_size) << total_items_count_string << delim
<< "time(ms)" << delim << scientific << setprecision(2) << total_executing_time;
for (auto item_times : total_interval_times)
{
cout << delim << item_times.first << "(ms)" << delim << item_times.second;
}
cout << "\n";
}
// Print time and memory consumed in ::call function
cout << func_name << delim << " Backend compilation(ms)" << delim << time_compile << delim
<< "call(ms)" << delim << time_call << delim << "memory consumption compile(B)" << delim
<< mem_compilation_consumed << delim << "call(B)" << delim << mem_call_consumed << delim
<< "RSS(B)" << delim << mem_current << endl;
cout.flags(saved_stream_flags); // Restore stream configuration to leave it in original state
}
bool runtime::intelgpu::IntelGPUBackend::is_supported_property(const Property prop) const bool runtime::intelgpu::IntelGPUBackend::is_supported_property(const Property prop) const
{ {
if (prop == Property::memory_attach) if (prop == Property::memory_attach)
......
src/ngraph/runtime/intelgpu/intelgpu_backend.hpp
View file @ 013c2381
...@@ -20,7 +20,6 @@ ...@@ -20,7 +20,6 @@
#include <memory> #include <memory>
#include <CPP/engine.hpp> #include <CPP/engine.hpp>
#include <CPP/network.hpp>
#include "ngraph/runtime/backend.hpp" #include "ngraph/runtime/backend.hpp"
...@@ -31,7 +30,6 @@ namespace ngraph ...@@ -31,7 +30,6 @@ namespace ngraph
namespace intelgpu namespace intelgpu
{ {
class IntelGPUBackend; class IntelGPUBackend;
class IntelGPUExecutable;
} }
} }
} }
...@@ -67,39 +65,3 @@ private: ...@@ -67,39 +65,3 @@ private:
bool m_disable_backend_optimizations = false; bool m_disable_backend_optimizations = false;
std::string m_cldnn_dump_dir = std::string("intelgpu_codegen"); std::string m_cldnn_dump_dir = std::string("intelgpu_codegen");
}; };
class ngraph::runtime::intelgpu::IntelGPUExecutable : public runtime::Executable
{
public:
IntelGPUExecutable(std::shared_ptr<Function> func,
std::shared_ptr<cldnn::network> network,
bool enable_timing,
bool enable_profile,
double compilation_time,
double consumed_memory,
size_t profile_lines_limit_count);
bool call(const std::vector<std::shared_ptr<runtime::Tensor>>& outputs,
const std::vector<std::shared_ptr<runtime::Tensor>>& inputs) override;
std::vector<PerformanceCounter> get_performance_data() const override;
private:
std::shared_ptr<Function> m_function;
std::shared_ptr<cldnn::network> m_cldnn_network = nullptr;
bool m_performance_counters_enabled = false;
bool m_profile_enable = false;
double m_compilation_time = 0.0;
double m_consumed_memory = 0.0;
long m_profile_lines_limit_count = 10;
std::string delim = std::string(":");
// Statistic related things
void print_call_performance(const std::shared_ptr<cldnn::network> network,
const std::shared_ptr<Function> func,
double time_compile,
double time_call,
double mem_compilation_consumed,
double mem_call_consumed,
double mem_current) const;
};
src/ngraph/runtime/intelgpu/intelgpu_executable.cpp 0 → 100644
View file @ 013c2381
//*****************************************************************************
// Copyright 2017-2019 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//*****************************************************************************
#include <iomanip>
#include "ngraph/runtime/intelgpu/intelgpu_executable.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.hpp"
#include "ngraph/runtime/intelgpu/intelgpu_tensor_view.hpp"
#include "ngraph/util.hpp"
using namespace std;
using namespace ngraph;
static void
memory_size_check(size_t memory_size, const shared_ptr<Node>& node, const string& function_name)
{
const size_t tensor_size = shape_size(node->get_shape()) * node->get_element_type().size();
if (memory_size != tensor_size)
{
ostringstream os;
os << "IntelGPU backend failed memory check. In \"" << function_name << "\" with Node \""
<< node->get_name() << "\" and " << node->get_shape() << " mismatched memory sizes "
<< tensor_size << " and " << memory_size;
throw invalid_argument(os.str());
}
}
static const string& get_input_name(const shared_ptr<Node>& op, size_t num = 0)
{
return op->get_inputs().at(num).get_tensor().get_name();
}
// The cldnn::network contains something like "generic_layer_0_Parameter_254_0" names
// This function should return "Parameter_254" from the example above
static string convert_cldnn_names(shared_ptr<Function> func, const string& cldnn_name)
{
const string key("_");
string result;
const size_t last_key = cldnn_name.rfind(key);
const size_t pre_last_key = cldnn_name.rfind(key, last_key - 1);
const size_t pre_pre_last_key = cldnn_name.rfind(key, pre_last_key - 1);
if (pre_pre_last_key == std::string::npos)
{
result = cldnn_name.substr(0, last_key);
}
else
{
result = cldnn_name.substr(pre_pre_last_key + 1, last_key - pre_pre_last_key - 1);
}
return result;
}
runtime::intelgpu::IntelGPUExecutable::IntelGPUExecutable(shared_ptr<Function> func,
shared_ptr<cldnn::network> network,
bool enable_timing,
bool enable_profile,
double compilation_time,
double consumed_memory,
size_t profile_lines_limit_count)
{
m_function = func;
m_cldnn_network = network;
m_performance_counters_enabled = enable_timing;
m_profile_enable = enable_profile;
m_compilation_time = compilation_time;
m_consumed_memory = consumed_memory;
m_profile_lines_limit_count = profile_lines_limit_count;
set_parameters_and_results(*func);
}
bool runtime::intelgpu::IntelGPUExecutable::call(const vector<shared_ptr<runtime::Tensor>>& outputs,
const vector<shared_ptr<runtime::Tensor>>& inputs)
{
double mem_call_consumed = 0.0f;
stopwatch timer_call;
if (m_cldnn_network == nullptr)
{
throw runtime_error("compile() must be called before call().");
}
if (m_profile_enable)
{
mem_call_consumed = runtime::intelgpu::get_max_memory_rss();
timer_call.start();
}
// Process input parameters. Correctness of parameters was validated by validate_call.
// Since we have no correlation between Function::m_parameters and inputs, there is
// we try to match them by index number in vectors.
for (size_t i = 0; i < inputs.size(); i++)
{
shared_ptr<runtime::intelgpu::IntelGPUTensorView> tv =
static_pointer_cast<runtime::intelgpu::IntelGPUTensorView>(inputs[i]);
const ParameterVector& input_params = get_parameters();
const string& tensor_name = input_params[i]->get_output_tensor().get_name();
m_cldnn_network->set_input_data(tensor_name, *tv->get_data_ptr());
}
// Execute network
map<cldnn::primitive_id, cldnn::network_output> result = m_cldnn_network->execute();
// Process output parameters. Correctness of parameters was validated by validate_call.
// Since we have no correlation between Function::m_results and outputs, there is
// we try to match them by index number in vectors.
for (size_t i = 0; i < m_function->get_output_size(); i++)
{
const shared_ptr<Node>& dst_node = m_function->get_output_op(i);
const size_t dst_shape_size = shape_size(dst_node->get_shape());
// We should not touch destination memory if it is not existed
if (!dst_shape_size)
{
continue;
}
shared_ptr<runtime::intelgpu::IntelGPUTensorView> ngraph_res =
static_pointer_cast<runtime::intelgpu::IntelGPUTensorView>(outputs[i]);
const string& tensor_name = get_input_name(dst_node);
auto result_memory = result.at(tensor_name).get_memory().pointer<char>();
memory_size_check(result_memory.size(), dst_node, m_function->get_name());
ngraph_res->write(result_memory.data(), 0, result_memory.size());
}
if (m_profile_enable)
{
timer_call.stop();
mem_call_consumed = runtime::intelgpu::get_max_memory_rss() - mem_call_consumed;
print_call_performance(m_cldnn_network,
m_function,
m_compilation_time,
timer_call.get_milliseconds(),
m_consumed_memory,
mem_call_consumed,
runtime::intelgpu::get_max_memory_rss());
// Output compile time only once
m_compilation_time = 0.0;
m_consumed_memory = 0.0;
}
return true;
}
vector<runtime::PerformanceCounter>
runtime::intelgpu::IntelGPUExecutable::get_performance_data() const
{
vector<runtime::PerformanceCounter> rc;
if (m_cldnn_network != nullptr && m_performance_counters_enabled)
{
const map<cldnn::primitive_id, cldnn::event>& primitives =
m_cldnn_network->get_executed_primitives();
for (const auto& p : primitives)
{
// Let's generate the primitive name that matches to the name in Function
const string primitive_name = convert_cldnn_names(m_function, p.first);
size_t usec = 0;
for (const auto& q : p.second.get_profiling_info())
{
if (q.name == string("executing"))
{
usec += chrono::duration_cast<
chrono::duration<size_t, chrono::milliseconds::period>>(
q.value->value())
.count();
}
}
const runtime::PerformanceCounter perf_counter(primitive_name.c_str(), usec, 1);
rc.push_back(perf_counter);
}
}
return rc;
}
static Node* get_node_by_name(const shared_ptr<Function> func, const string& name)
{
for (shared_ptr<Node> node : func->get_ops())
{
if (node->get_name() == name)
{
return node.get();
}
}
return nullptr;
}
void runtime::intelgpu::IntelGPUExecutable::print_call_performance(
const shared_ptr<cldnn::network> network,
const shared_ptr<Function> func,
double time_compile,
double time_call,
double mem_compilation_consumed,
double mem_call_consumed,
double mem_current) const
{
struct data_item
{
string item_name;
map<string, double> item_times;
};
const string& func_name = func->get_name();
const map<cldnn::primitive_id, cldnn::event>& primitives = network->get_executed_primitives();
size_t limit_count = m_profile_lines_limit_count;
multimap<double, data_item> data;
map<string, double> total_interval_times;
double total_executing_time = 0;
size_t total_items_count = 0;
size_t max_item_name_size = 0;
ios_base::fmtflags saved_stream_flags(cout.flags()); // Save stream flags to restore them later
if (m_profile_lines_limit_count > 0)
{
// Extract profiling statistic, calculate summary and sort
for (auto& prim : primitives)
{
double executing_time = 0;
data_item item;
item.item_name = prim.first;
max_item_name_size = max(max_item_name_size, prim.first.size());
for (auto& prof_info : prim.second.get_profiling_info())
{
const string& interval_name = prof_info.name;
double interval =
chrono::duration_cast<chrono::duration<double, chrono::milliseconds::period>>(
prof_info.value->value())
.count();
item.item_times[interval_name] = interval;
// Get the Key time to sort by
if (interval_name == "executing")
{
executing_time += interval;
}
// Accumulate total time for each interval
if (total_interval_times.find(interval_name) == total_interval_times.end())
{
total_interval_times[interval_name] = interval;
}
else
{
total_interval_times[interval_name] += interval;
}
}
data.emplace(executing_time, item);
total_executing_time += executing_time;
++total_items_count;
}
// Print statistic for each primitive in the cldnn::network
for (auto it = data.rbegin(); (it != data.rend()) && (limit_count > 0); ++it, --limit_count)
{
const string ngraph_node_name = convert_cldnn_names(func, it->second.item_name);
const Node* ngraph_node = get_node_by_name(func, ngraph_node_name);
cout << func_name << delim << setw(max_item_name_size) << it->second.item_name << delim
<< "time(ms)" << delim << scientific << setprecision(2) << it->first;
for (auto item : it->second.item_times)
{
cout << delim << item.first << "(ms)" << delim << item.second;
}
cout << delim << ngraph_node_name;
if (ngraph_node) // it might be initialized by nullptr
{
// print all input shapes for the Node
size_t arg_idx = 0;
for (const descriptor::Input& op_input : ngraph_node->get_inputs())
{
cout << delim << op_input.get_element_type().c_type_string() << " input"
<< arg_idx << vector_to_string(op_input.get_shape());
++arg_idx;
}
// print all output shapes for the Node
arg_idx = 0;
for (const descriptor::Output& op_output : ngraph_node->get_outputs())
{
cout << delim << op_output.get_element_type().c_type_string() << " output"
<< arg_idx << vector_to_string(op_output.get_shape());
++arg_idx;
}
}
cout << "\n";
}
// Print bottom line summary
const string total_items_count_string = "Total(cldnn " + to_string(total_items_count) +
", ngraph " + to_string(func->get_ops().size()) +
")";
cout << func_name << delim << setw(max_item_name_size) << total_items_count_string << delim
<< "time(ms)" << delim << scientific << setprecision(2) << total_executing_time;
for (auto item_times : total_interval_times)
{
cout << delim << item_times.first << "(ms)" << delim << item_times.second;
}
cout << "\n";
}
// Print time and memory consumed in ::call function
cout << func_name << delim << " Backend compilation(ms)" << delim << time_compile << delim
<< "call(ms)" << delim << time_call << delim << "memory consumption compile(B)" << delim
<< mem_compilation_consumed << delim << "call(B)" << delim << mem_call_consumed << delim
<< "RSS(B)" << delim << mem_current << endl;
cout.flags(saved_stream_flags); // Restore stream configuration to leave it in original state
}
src/ngraph/runtime/intelgpu/intelgpu_executable.hpp 0 → 100644
View file @ 013c2381
//*****************************************************************************
// Copyright 2017-2019 Intel Corporation
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//*****************************************************************************
#pragma once
#include <CPP/network.hpp>
#include "ngraph/runtime/tensor.hpp"
namespace ngraph
{
namespace runtime
{
namespace intelgpu
{
class IntelGPUExecutable;
}
}
}
class ngraph::runtime::intelgpu::IntelGPUExecutable : public runtime::Executable
{
public:
IntelGPUExecutable(std::shared_ptr<Function> func,
std::shared_ptr<cldnn::network> network,
bool enable_timing,
bool enable_profile,
double compilation_time,
double consumed_memory,
size_t profile_lines_limit_count);
bool call(const std::vector<std::shared_ptr<runtime::Tensor>>& outputs,
const std::vector<std::shared_ptr<runtime::Tensor>>& inputs) override;
std::vector<PerformanceCounter> get_performance_data() const override;
private:
std::shared_ptr<Function> m_function;
std::shared_ptr<cldnn::network> m_cldnn_network = nullptr;
bool m_performance_counters_enabled = false;
bool m_profile_enable = false;
double m_compilation_time = 0.0;
double m_consumed_memory = 0.0;
long m_profile_lines_limit_count = 10;
std::string delim = std::string(":");
// Statistic related things
void print_call_performance(const std::shared_ptr<cldnn::network> network,
const std::shared_ptr<Function> func,
double time_compile,
double time_call,
double mem_compilation_consumed,
double mem_call_consumed,
double mem_current) const;
};
src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.cpp
View file @ 013c2381
...@@ -14,6 +14,9 @@ ...@@ -14,6 +14,9 @@
// limitations under the License. // limitations under the License.
//***************************************************************************** //*****************************************************************************
#include <sys/resource.h>
#include <sys/time.h>
#include <CPP/concatenation.hpp> #include <CPP/concatenation.hpp>
#include <CPP/custom_gpu_primitive.hpp> #include <CPP/custom_gpu_primitive.hpp>
#include <CPP/reshape.hpp> #include <CPP/reshape.hpp>
...@@ -1515,3 +1518,19 @@ void runtime::intelgpu::do_reshape_operation(cldnn::topology& topology, ...@@ -1515,3 +1518,19 @@ void runtime::intelgpu::do_reshape_operation(cldnn::topology& topology,
{1}); {1});
topology.add(op_reshape); topology.add(op_reshape);
} }
size_t runtime::intelgpu::get_max_memory_rss()
{
size_t result = 0;
struct rusage usage;
if (getrusage(RUSAGE_SELF, &usage) == 0)
{
result = usage.ru_maxrss; // the value is in kilobytes
// aligne result to return bytes
result *= 1000;
}
return result;
}
src/ngraph/runtime/intelgpu/intelgpu_op_custom_kernels.hpp
View file @ 013c2381
...@@ -33,6 +33,8 @@ namespace ngraph ...@@ -33,6 +33,8 @@ namespace ngraph
{ {
namespace intelgpu namespace intelgpu
{ {
size_t get_max_memory_rss();
void do_pad_operation(cldnn::topology& topology, void do_pad_operation(cldnn::topology& topology,
const std::string& input_name, const std::string& input_name,
const Shape& input_shape, const Shape& input_shape,
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment