Add more logs

Author: nikita-kud

samples/cpp/benchmark_app/main.cpp

@@ -37,6 +37,42 @@
 #include "utils.hpp"
 // clang-format on
 
+#if defined _WIN32
+
+#include <windows.h>
+
+#include <psapi.h>
+
+int64_t getPeakMemoryUsage() {
+    PROCESS_MEMORY_COUNTERS memCounters;
+    GetProcessMemoryInfo(GetCurrentProcess(), &memCounters, sizeof(memCounters));
+    return memCounters.PeakWorkingSetSize / 1000;
+}
+
+#else
+
+#include <fstream>
+#include <regex>
+#include <sstream>
+
+int64_t getPeakMemoryUsage() {
+    size_t peakMemUsageKB = 0;
+
+    std::ifstream statusFile("/proc/self/status");
+    std::string line;
+    std::regex vmPeakRegex("VmPeak:");
+    std::smatch vmMatch;
+    while (std::getline(statusFile, line)) {
+        if (std::regex_search(line, vmMatch, vmPeakRegex)) {
+            std::istringstream iss(vmMatch.suffix());
+            iss >> peakMemUsageKB;
+        }
+    }
+    return static_cast<int64_t>(peakMemUsageKB);
+}
+
+#endif
+
 namespace {
 bool parse_and_check_command_line(int argc, char* argv[]) {
     // ---------------------------Parsing and validating input
@@ -775,7 +811,16 @@ int main(int argc, char* argv[]) {
             if (!modelStream.is_open()) {
                 throw std::runtime_error("Cannot open model file " + FLAGS_m);
             }
+            
+            auto importModelBegin = std::chrono::steady_clock::now();
+
             compiledModel = core.import_model(modelStream, device_name, device_config);
+
+            auto importModelEnd = std::chrono::steady_clock::now();
+            std::cout << "Import model time: "
+                << std::chrono::duration_cast<std::chrono::milliseconds>(importModelEnd - importModelBegin).count() << " ms"
+                << std::endl;
+
             modelStream.close();
 
             auto duration_ms = get_duration_ms_till_now(startTime);

src/plugins/intel_npu/src/compiler_adapter/src/plugin_compiler_adapter.cpp

@@ -110,6 +110,8 @@ std::vector<std::shared_ptr<IGraph>> PluginCompilerAdapter::compileWS(const std:
                                                                       const Config& config) const {
     OV_ITT_TASK_CHAIN(COMPILE_BLOB, itt::domains::NPUPlugin, "PluginCompilerAdapter", "compileWS");
 
+    auto compileNetBegin = std::chrono::steady_clock::now();
+
     std::shared_ptr<NetworkDescription> initNetworkDescription;
     std::shared_ptr<NetworkDescription> mainNetworkDescription;
 
@@ -188,6 +190,11 @@ std::vector<std::shared_ptr<IGraph>> PluginCompilerAdapter::compileWS(const std:
 
     _logger.debug("compile end");
 
+    auto compileNetEnd = std::chrono::steady_clock::now();
+    std::cout << "Compile net time: "
+              << std::chrono::duration_cast<std::chrono::milliseconds>(compileNetEnd - compileNetBegin).count() << " ms"
+              << std::endl;
+
     ze_graph_handle_t initGraphHandle = nullptr;
     ze_graph_handle_t mainGraphHandle = nullptr;
 

src/plugins/intel_npu/src/compiler_adapter/src/plugin_graph.cpp

@@ -67,7 +67,9 @@ void PluginGraph::custom_export(std::ostream& stream,
     } else {
         if (_logger.level() >= ov::log::Level::INFO) {
             std::stringstream str;
-            str << "Blob size: " << _blob.size() + initBlob.size() + 4 * sizeof(size_t) + xmlSize + binSize
+            str << "Blob size: " << _blob.size() + initBlob.size()
+                << std::endl;
+            str << "Blob size with weights: " << _blob.size() + initBlob.size() + 4 * sizeof(size_t) + xmlSize + binSize
                 << std::endl;
             _logger.info(str.str().c_str());
         }