add --client_package_build option

cmake/CMakeLists.txt

@@ -151,6 +151,7 @@ option(onnxruntime_DISABLE_SPARSE_TENSORS "Disable sparse tensors data types" OF
 option(onnxruntime_DISABLE_OPTIONAL_TYPE "Disable optional type" OFF)
 option(onnxruntime_DISABLE_FLOAT8_TYPES "Disable float 8 types" OFF)
 option(onnxruntime_MINIMAL_BUILD "Exclude as much as possible from the build. Support ORT format models. No support for ONNX format models." OFF)
+option(onnxruntime_CLIENT_PACKAGE_BUILD "Enables default settings that are more appropriate for client/on-device workloads." OFF)
 cmake_dependent_option(onnxruntime_DISABLE_RTTI "Disable RTTI" ON "NOT onnxruntime_ENABLE_PYTHON;NOT onnxruntime_USE_CUDA" OFF)
 # For now onnxruntime_DISABLE_EXCEPTIONS will only work with onnxruntime_MINIMAL_BUILD, more changes (ONNX, non-CPU EP, ...) are required to run this standalone
 cmake_dependent_option(onnxruntime_DISABLE_EXCEPTIONS "Disable exception handling. Requires onnxruntime_MINIMAL_BUILD currently." ON "onnxruntime_MINIMAL_BUILD;NOT onnxruntime_ENABLE_PYTHON" OFF)

cmake/adjust_global_compile_flags.cmake

@@ -95,6 +95,11 @@ if (onnxruntime_MINIMAL_BUILD)
   endif()
 endif()
 
+# ORT build with default settings more appropriate for client/on-device workloads.
+if (onnxruntime_CLIENT_PACKAGE_BUILD)
+  add_compile_definitions(ORT_CLIENT_PACKAGE_BUILD)
+endif()
+
 if (onnxruntime_ENABLE_LTO)
     include(CheckIPOSupported)
     check_ipo_supported(RESULT ipo_enabled OUTPUT ipo_output)

include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h

@@ -148,7 +148,9 @@ static const char* const kOrtSessionOptionsUseDeviceAllocatorForInitializers = "
 
 // Configure whether to allow the inter_op/intra_op threads spinning a number of times before blocking
 // "0": thread will block if found no job to run
-// "1": default, thread will spin a number of times before blocking
+// "1": thread will spin a number of times before blocking
+// The default is "0" when ORT is built with "ORT_CLIENT_PACKAGE_BUILD" and "1" otherwise.
+// Thread spinning is disabled by default for client/on-device workloads to reduce cpu utilization and improve power efficiency.
 static const char* const kOrtSessionOptionsConfigAllowInterOpSpinning = "session.inter_op.allow_spinning";
 static const char* const kOrtSessionOptionsConfigAllowIntraOpSpinning = "session.intra_op.allow_spinning";
 

onnxruntime/core/session/inference_session.cc

@@ -423,7 +423,13 @@ void InferenceSession::ConstructorCommon(const SessionOptions& session_options,
     {
       if (!external_intra_op_thread_pool_) {
         bool allow_intra_op_spinning =
+#if !defined(ORT_CLIENT_PACKAGE_BUILD)
             session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigAllowIntraOpSpinning, "1") == "1";
+#else
+            // default KOrtSessionOptionsConfigAllowIntraOpSpinning to "0" for ORT builds targeting client/on-device workloads,
+            // to reduce CPU utilization and improve power efficiency.
+            session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigAllowIntraOpSpinning, "0") == "1";
+#endif
         OrtThreadPoolParams to = session_options_.intra_op_param;
         std::basic_stringstream<ORTCHAR_T> ss;
         if (to.name) {
@@ -461,7 +467,13 @@ void InferenceSession::ConstructorCommon(const SessionOptions& session_options,
     if (session_options_.execution_mode == ExecutionMode::ORT_PARALLEL) {
       if (!external_inter_op_thread_pool_) {
         bool allow_inter_op_spinning =
+#if !defined(ORT_CLIENT_PACKAGE_BUILD)
             session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigAllowInterOpSpinning, "1") == "1";
+#else
+            // default kOrtSessionOptionsConfigAllowInterOpSpinning to "0" for ORT builds targeting client/on-device workloads,
+            // to reduce CPU utilization and improve power efficiency.
+            session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigAllowInterOpSpinning, "0") == "1";
+#endif
         OrtThreadPoolParams to = session_options_.inter_op_param;
         to.auto_set_affinity = to.thread_pool_size == 0 && session_options_.execution_mode == ExecutionMode::ORT_SEQUENTIAL;
         std::basic_stringstream<ORTCHAR_T> ss;

onnxruntime/core/util/thread_utils.h

@@ -19,7 +19,13 @@ struct OrtThreadPoolParams {
   bool auto_set_affinity = false;
 
   // If it is true, the thread pool will spin a while after the queue became empty.
+#if !defined(ORT_CLIENT_PACKAGE_BUILD)
   bool allow_spinning = true;
+#else
+  // default allow_spinning to false for ORT builds targeting client/on-device workloads,
+  // to reduce CPU utilization and improve power efficiency.
+  bool allow_spinning = false;
+#endif
 
   // It it is non-negative, thread pool will split a task by a decreasing block size
   // of remaining_of_total_iterations / (num_of_threads * dynamic_block_base_)

tools/ci_build/build.py

@@ -470,6 +470,7 @@ def generate_build_tree(
             else "OFF"
         ),
         "-Donnxruntime_REDUCED_OPS_BUILD=" + ("ON" if is_reduced_ops_build(args) else "OFF"),
+        "-Donnxruntime_CLIENT_PACKAGE_BUILD=" + ("ON" if args.client_package_build else "OFF"),
         "-Donnxruntime_BUILD_MS_EXPERIMENTAL_OPS=" + ("ON" if args.ms_experimental else "OFF"),
         "-Donnxruntime_ENABLE_LTO=" + ("ON" if args.enable_lto else "OFF"),
         "-Donnxruntime_USE_ACL=" + ("ON" if args.use_acl else "OFF"),

tools/ci_build/build_args.py

@@ -527,6 +527,15 @@ def add_size_reduction_args(parser: argparse.ArgumentParser) -> None:
     )
 
 
+def add_client_package_args(parser: argparse.ArgumentParser) -> None:
+    """Adds arguments for client package build package."""
+    parser.add_argument(
+        "--client_package_build",
+        action="store_true",
+        help="Create ORT package with default settings more appropriate for client/on-device workloads.",
+    )
+
+
 def add_python_binding_args(parser: argparse.ArgumentParser) -> None:
     """Adds arguments for Python bindings."""
     parser.add_argument("--enable_pybind", action="store_true", help="Enable Python bindings.")
@@ -833,6 +842,7 @@ def convert_arg_line_to_args(self, arg_line: str) -> list[str]:  # Use list[str]
     add_dependency_args(parser)
     add_extension_args(parser)
     add_size_reduction_args(parser)
+    add_client_package_args(parser)
 
     # Language Bindings
     add_python_binding_args(parser)

tools/ci_build/github/azure-pipelines/templates/qnn-ep-win.yml

@@ -20,7 +20,7 @@ stages:
       name: ${{ parameters.qnn_ep_build_pool_name }}
     variables:
       OrtPackageId: ${{ parameters.OrtNugetPackageId }}
-      commonBuildArgs: '--compile_no_warning_as_error --skip_submodule_sync --build_shared_lib --cmake_generator "Visual Studio 17 2022" --config ${{ parameters.build_config }} --parallel --use_vcpkg --use_vcpkg_ms_internal_asset_cache --use_binskim_compliant_compile_flags '
+      commonBuildArgs: '--compile_no_warning_as_error --skip_submodule_sync --build_shared_lib --client_package_build --cmake_generator "Visual Studio 17 2022" --config ${{ parameters.build_config }} --parallel --use_vcpkg --use_vcpkg_ms_internal_asset_cache --use_binskim_compliant_compile_flags '
 
     steps:
     - template: set-version-number-variables-step.yml

tools/ci_build/github/azure-pipelines/win-qnn-arm64-ci-pipeline.yml

@@ -50,7 +50,7 @@ jobs:
     matrix:
       SHARED_LIB:
         QnnLibKind: 'shared_lib'
-        ExtraQnnBuildArgs: ''
+        ExtraQnnBuildArgs: '--client_package_build'
       STATIC_LIB:
         QnnLibKind: 'static_lib'
         ExtraQnnBuildArgs: ''