Debug logs

Author: eshiryae

samples/cpp/whisper_speech_recognition/whisper_speech_recognition.cpp

@@ -5,15 +5,20 @@
 #include "openvino/genai/whisper_pipeline.hpp"
 
 int main(int argc, char* argv[]) try {
-    if (argc < 3 || argc > 4) {
+    if (argc < 3 || argc > 5) {
         throw std::runtime_error(std::string{"Usage: "} + argv[0] + " <MODEL_DIR> \"<WAV_FILE_PATH>\" <DEVICE>");
     }
 
     std::filesystem::path models_path = argv[1];
     std::string wav_file_path = argv[2];
-    std::string device = (argc == 4) ? argv[3] : "CPU";  // Default to CPU if no device is provided
+    std::string device = (argc >= 4) ? argv[3] : "CPU";  // Default to CPU if no device is provided
 
-    ov::genai::WhisperPipeline pipeline(models_path, device);
+    ov::AnyMap npu_config = {};
+    if (device == "NPU") 
+        npu_config = {{"NPU_USE_NPUW", "YES"}, {"NPUW_DEVICES", "CPU"}, {"NPUW_ONLINE_PIPELINE", "NONE"}};
+    if (device == "NPU" && argc == 5 && argv[4] == "STATIC")
+        npu_config = {{"STATIC_PIPELINE", "YES"}};
+    ov::genai::WhisperPipeline pipeline(models_path, device, npu_config);
 
     ov::genai::WhisperGenerationConfig config = pipeline.get_generation_config();
     // 'task' and 'language' parameters are supported for multilingual models only
@@ -32,6 +37,15 @@ int main(int argc, char* argv[]) try {
         std::cout << "timestamps: [" << chunk.start_ts << ", " << chunk.end_ts << "] text: " << chunk.text << "\n";
     }
 
+    std::cout << std::fixed << std::setprecision(2) << "\n\nLoad time: " << result.perf_metrics.get_load_time() << " ms" << std::endl;
+    std::cout << std::fixed << std::setprecision(2) << "Generate time: " << result.perf_metrics.get_generate_duration().mean << " +/- " << result.perf_metrics.get_generate_duration().std << " ms" << std::endl;
+    std::cout << std::fixed << std::setprecision(2) << "Tokenization time: " << result.perf_metrics.get_tokenization_duration().mean << " +/- " << result.perf_metrics.get_tokenization_duration().std << " ms" << std::endl;
+    std::cout << std::fixed << std::setprecision(2) << "Detokenization time: " << result.perf_metrics.get_detokenization_duration().mean << " +/- " << result.perf_metrics.get_detokenization_duration().std << " ms" << std::endl;
+    std::cout << std::fixed << std::setprecision(2) << "TTFT: " << result.perf_metrics.get_ttft().mean << " +/- " << result.perf_metrics.get_ttft().std << " ms" << std::endl;
+    std::cout << std::fixed << std::setprecision(2) << "TPOT: " << result.perf_metrics.get_tpot().mean << " +/- " << result.perf_metrics.get_tpot().std << " ms" << std::endl;
+    std::cout << std::fixed << std::setprecision(2) << "Throughput : " << result.perf_metrics.get_throughput().mean << " +/- " << result.perf_metrics.get_throughput().std << " tokens/s" << std::endl;
+
+
 } catch (const std::exception& error) {
     try {
         std::cerr << error.what() << '\n';

src/cpp/src/whisper/pipeline.cpp

@@ -71,6 +71,9 @@ class WhisperPipeline::WhisperPipelineStatefulImpl : public WhisperPipeline::Whi
                                 const ov::AnyMap& properties)
         : WhisperPipelineImplBase{models_path},
           m_sampler(m_tokenizer) {
+
+        std::cout << "[DEBUG] Running WhisperPipelineStatefulImpl ..." << std::endl;
+
         ov::Core core = utils::singleton_core();
         ov::CompiledModel compiled_model;
         if (device == "NPU") {

src/cpp/src/whisper/pipeline_static.cpp

@@ -947,6 +947,8 @@ WhisperPipeline::StaticWhisperPipeline::StaticWhisperPipeline(const std::filesys
     , m_sampler(m_tokenizer) {
     ov::Core core = utils::singleton_core();
 
+    std::cout << "[DEBUG] Running StaticWhisperPipeline ..." << std::endl;
+
     // Remove "STATIC_PIPELINE" as we don't need to pass it further
     auto model_properties = properties;
     utils::pop_option(model_properties, "STATIC_PIPELINE");