[SYCL] Optimize getSyclObjImpl return values (#17491)

Use references as local variables to assign the return value of
getSyclObjImpl.

Author: slawekptak

sycl/source/backend.cpp

@@ -319,7 +319,7 @@ kernel make_kernel(const context &TargetContext,
                    backend Backend) {
   const auto &Adapter = getAdapter(Backend);
   const auto &ContextImpl = getSyclObjImpl(TargetContext);
-  const auto KernelBundleImpl = getSyclObjImpl(KernelBundle);
+  const auto &KernelBundleImpl = getSyclObjImpl(KernelBundle);
 
   // For Level-Zero expect exactly one device image in the bundle. This is
   // natural for interop kernel to get created out of a single native

sycl/source/detail/image_impl.cpp

@@ -259,7 +259,7 @@ image_channel_type convertChannelType(ur_image_channel_type_t Type) {
 }
 
 template <typename T>
-static void getImageInfo(const ContextImplPtr Context, ur_image_info_t Info,
+static void getImageInfo(const ContextImplPtr &Context, ur_image_info_t Info,
                          T &Dest, ur_mem_handle_t InteropMemObject) {
   const AdapterPtr &Adapter = Context->getAdapter();
   Adapter->call<UrApiKind::urMemImageGetInfo>(InteropMemObject, Info, sizeof(T),
@@ -274,7 +274,7 @@ image_impl::image_impl(cl_mem MemObject, const context &SyclContext,
             std::move(Allocator)),
       MDimensions(Dimensions), MRange({0, 0, 0}) {
   ur_mem_handle_t Mem = ur::cast<ur_mem_handle_t>(BaseT::MInteropMemObject);
-  const ContextImplPtr Context = getSyclObjImpl(SyclContext);
+  const ContextImplPtr &Context = getSyclObjImpl(SyclContext);
   const AdapterPtr &Adapter = Context->getAdapter();
   Adapter->call<UrApiKind::urMemGetInfo>(Mem, UR_MEM_INFO_SIZE, sizeof(size_t),
                                          &(BaseT::MSizeInBytes), nullptr);

sycl/source/detail/kernel_bundle_impl.hpp

@@ -255,7 +255,7 @@ class kernel_bundle_impl {
     removeDuplicateImages();
 
     for (const kernel_bundle<bundle_state::object> &Bundle : ObjectBundles) {
-      const KernelBundleImplPtr BundlePtr = getSyclObjImpl(Bundle);
+      const KernelBundleImplPtr &BundlePtr = getSyclObjImpl(Bundle);
       for (const std::pair<const std::string, std::vector<unsigned char>>
                &SpecConst : BundlePtr->MSpecConstValues) {
         MSpecConstValues[SpecConst.first] = SpecConst.second;
@@ -449,7 +449,7 @@ class kernel_bundle_impl {
                                   const std::string &SourceStr,
                                   ur_program_handle_t &UrProgram) {
     using ContextImplPtr = std::shared_ptr<sycl::detail::context_impl>;
-    ContextImplPtr ContextImpl = getSyclObjImpl(MContext);
+    const ContextImplPtr &ContextImpl = getSyclObjImpl(MContext);
     const AdapterPtr &Adapter = ContextImpl->getAdapter();
 
     std::string UserArgs = syclex::detail::userArgsAsString(BuildOptions);
@@ -494,13 +494,13 @@ class kernel_bundle_impl {
            "bundle_state::ext_oneapi_source required");
 
     using ContextImplPtr = std::shared_ptr<sycl::detail::context_impl>;
-    ContextImplPtr ContextImpl = getSyclObjImpl(MContext);
+    const ContextImplPtr &ContextImpl = getSyclObjImpl(MContext);
     const AdapterPtr &Adapter = ContextImpl->getAdapter();
 
     std::vector<ur_device_handle_t> DeviceVec;
     DeviceVec.reserve(Devices.size());
     for (const auto &SyclDev : Devices) {
-      DeviceImplPtr DevImpl = getSyclObjImpl(SyclDev);
+      const DeviceImplPtr &DevImpl = getSyclObjImpl(SyclDev);
       if (!ContextImpl->hasDevice(DevImpl)) {
         throw sycl::exception(make_error_code(errc::invalid),
                               "device not part of kernel_bundle context");
@@ -785,7 +785,7 @@ class kernel_bundle_impl {
     std::transform(MDeviceGlobalNames.begin(), MDeviceGlobalNames.end(),
                    std::back_inserter(DeviceGlobalIDs),
                    [&](const std::string &DGName) { return MPrefix + DGName; });
-    auto ContextImpl = getSyclObjImpl(MContext);
+    const auto &ContextImpl = getSyclObjImpl(MContext);
     for (DeviceGlobalMapEntry *Entry :
          ProgramManager::getInstance().getDeviceGlobalEntries(
              DeviceGlobalIDs)) {
@@ -840,7 +840,7 @@ class kernel_bundle_impl {
     const std::shared_ptr<detail::device_image_impl> &DeviceImageImpl =
         detail::getSyclObjImpl(MDeviceImages[0].getMain());
     ur_program_handle_t UrProgram = DeviceImageImpl->get_ur_program_ref();
-    ContextImplPtr ContextImpl = getSyclObjImpl(MContext);
+    const ContextImplPtr &ContextImpl = getSyclObjImpl(MContext);
     const AdapterPtr &Adapter = ContextImpl->getAdapter();
     ur_kernel_handle_t UrKernel = nullptr;
     Adapter->call<UrApiKind::urKernelCreate>(UrProgram, AdjustedName.c_str(),

sycl/source/detail/program_manager/program_manager.cpp

@@ -233,7 +233,7 @@ ProgramManager::createURProgram(const RTDeviceBinaryImage &Img,
   const auto &ProgMetadata = Img.getProgramMetadataUR();
 
   // Load the image
-  const ContextImplPtr Ctx = getSyclObjImpl(Context);
+  const ContextImplPtr &Ctx = getSyclObjImpl(Context);
   std::vector<const uint8_t *> Binaries(
       Devices.size(), const_cast<uint8_t *>(RawImg.BinaryStart));
   std::vector<size_t> Lengths(Devices.size(), ImgSize);
@@ -2696,7 +2696,8 @@ void ProgramManager::bringSYCLDeviceImagesToState(
 
   for (DevImgPlainWithDeps &ImgWithDeps : DeviceImages) {
     device_image_plain &MainImg = ImgWithDeps.getMain();
-    const bundle_state DevImageState = getSyclObjImpl(MainImg)->get_state();
+    const DeviceImageImplPtr &MainImgImpl = getSyclObjImpl(MainImg);
+    const bundle_state DevImageState = MainImgImpl->get_state();
     // At this time, there is no circumstance where a device image should ever
     // be in the source state. That not good.
     assert(DevImageState != bundle_state::ext_oneapi_source);
@@ -2712,9 +2713,8 @@ void ProgramManager::bringSYCLDeviceImagesToState(
       break;
     case bundle_state::object:
       if (DevImageState == bundle_state::input) {
-        ImgWithDeps =
-            compile(ImgWithDeps, getSyclObjImpl(MainImg)->get_devices(),
-                    /*PropList=*/{});
+        ImgWithDeps = compile(ImgWithDeps, MainImgImpl->get_devices(),
+                              /*PropList=*/{});
         break;
       }
       // Device image is expected to be object state then.
@@ -2728,7 +2728,7 @@ void ProgramManager::bringSYCLDeviceImagesToState(
         assert(DevImageState != bundle_state::ext_oneapi_source);
         break;
       case bundle_state::input:
-        ImgWithDeps = build(ImgWithDeps, getSyclObjImpl(MainImg)->get_devices(),
+        ImgWithDeps = build(ImgWithDeps, MainImgImpl->get_devices(),
                             /*PropList=*/{});
         break;
       case bundle_state::object: {
@@ -2742,7 +2742,7 @@ void ProgramManager::bringSYCLDeviceImagesToState(
         break;
       }
       case bundle_state::executable:
-        ImgWithDeps = build(ImgWithDeps, getSyclObjImpl(MainImg)->get_devices(),
+        ImgWithDeps = build(ImgWithDeps, MainImgImpl->get_devices(),
                             /*PropList=*/{});
         break;
       }
@@ -2884,7 +2884,8 @@ static void mergeImageData(const std::vector<device_image_plain> &Imgs,
                            std::vector<unsigned char> &NewSpecConstBlob,
                            device_image_impl::SpecConstMapT &NewSpecConstMap) {
   for (const device_image_plain &Img : Imgs) {
-    std::shared_ptr<device_image_impl> DeviceImageImpl = getSyclObjImpl(Img);
+    const std::shared_ptr<device_image_impl> &DeviceImageImpl =
+        getSyclObjImpl(Img);
     // Duplicates are not expected here, otherwise urProgramLink should fail
     if (DeviceImageImpl->get_kernel_ids_ptr())
       KernelIDs.insert(KernelIDs.end(),
@@ -2946,16 +2947,15 @@ ProgramManager::link(const DevImgPlainWithDeps &ImgWithDeps,
   std::string LinkOptionsStr;
   applyLinkOptionsFromEnvironment(LinkOptionsStr);
   const device_image_plain &MainImg = ImgWithDeps.getMain();
+  const std::shared_ptr<device_image_impl> &InputImpl = getSyclObjImpl(MainImg);
   if (LinkOptionsStr.empty()) {
-    const std::shared_ptr<device_image_impl> &InputImpl =
-        getSyclObjImpl(MainImg);
     appendLinkOptionsFromImage(LinkOptionsStr,
                                *(InputImpl->get_bin_image_ref()));
   }
   // Should always come last!
   appendLinkEnvironmentVariablesThatAppend(LinkOptionsStr);
-  const context &Context = getSyclObjImpl(MainImg)->get_context();
-  const ContextImplPtr ContextImpl = getSyclObjImpl(Context);
+  const context &Context = InputImpl->get_context();
+  const ContextImplPtr &ContextImpl = getSyclObjImpl(Context);
   const AdapterPtr &Adapter = ContextImpl->getAdapter();
 
   ur_program_handle_t LinkedProg = nullptr;
@@ -2981,8 +2981,7 @@ ProgramManager::link(const DevImgPlainWithDeps &ImgWithDeps,
 
   if (Error != UR_RESULT_SUCCESS) {
     if (LinkedProg) {
-      const std::string ErrorMsg =
-          getProgramBuildLog(LinkedProg, std::move(ContextImpl));
+      const std::string ErrorMsg = getProgramBuildLog(LinkedProg, ContextImpl);
       throw sycl::exception(make_error_code(errc::build), ErrorMsg);
     }
     throw set_ur_error(exception(make_error_code(errc::build), "link() failed"),
@@ -3008,7 +3007,7 @@ ProgramManager::link(const DevImgPlainWithDeps &ImgWithDeps,
     }
   }
 
-  auto BinImg = getSyclObjImpl(MainImg)->get_bin_image_ref();
+  auto BinImg = InputImpl->get_bin_image_ref();
   DeviceImageImplPtr ExecutableImpl =
       std::make_shared<detail::device_image_impl>(
           BinImg, Context, Devs, bundle_state::executable, std::move(KernelIDs),
@@ -3037,7 +3036,6 @@ ProgramManager::build(const DevImgPlainWithDeps &DevImgWithDeps,
       getSyclObjImpl(DevImgWithDeps.getMain());
 
   const context Context = MainInputImpl->get_context();
-  const ContextImplPtr ContextImpl = getSyclObjImpl(Context);
 
   std::vector<const RTDeviceBinaryImage *> BinImgs;
   BinImgs.reserve(DevImgWithDeps.size());
@@ -3089,7 +3087,7 @@ ProgramManager::getOrCreateKernel(const context &Context,
         PropList, NoAllowedPropertiesCheck, NoAllowedPropertiesCheck);
   }
 
-  const ContextImplPtr Ctx = getSyclObjImpl(Context);
+  const ContextImplPtr &Ctx = getSyclObjImpl(Context);
 
   KernelProgramCache &Cache = Ctx->getKernelProgramCache();
 

sycl/source/detail/queue_impl.cpp

@@ -303,7 +303,7 @@ sycl::detail::optional<event> queue_impl::getLastEvent() {
 }
 
 void queue_impl::addEvent(const event &Event) {
-  EventImplPtr EImpl = getSyclObjImpl(Event);
+  const EventImplPtr &EImpl = getSyclObjImpl(Event);
   assert(EImpl && "Event implementation is missing");
   auto *Cmd = static_cast<Command *>(EImpl->getCommand());
   if (!Cmd) {

sycl/source/detail/queue_impl.hpp

@@ -848,7 +848,7 @@ class queue_impl {
       Handler.depends_on(Deps.LastBarrier);
 
     auto EventRet = Handler.finalize();
-    EventImplPtr EventRetImpl = getSyclObjImpl(EventRet);
+    const EventImplPtr &EventRetImpl = getSyclObjImpl(EventRet);
     if (Type == CGType::CodeplayHostTask)
       Deps.UnenqueuedCmdEvents.push_back(std::move(EventRetImpl));
     else if (Type == CGType::Barrier || Type == CGType::BarrierWaitlist) {

sycl/source/detail/ur.cpp

@@ -49,7 +49,7 @@ namespace pi {
 void contextSetExtendedDeleter(const sycl::context &context,
                                pi_context_extended_deleter func,
                                void *user_data) {
-  auto impl = getSyclObjImpl(context);
+  const auto &impl = getSyclObjImpl(context);
   const auto &Adapter = impl->getAdapter();
   Adapter->call<UrApiKind::urContextSetExtendedDeleter>(
       impl->getHandleRef(),

sycl/source/kernel_bundle.cpp

@@ -481,7 +481,8 @@ exe_kb build_from_source(
     LogPtr = &Log;
   std::vector<device> UniqueDevices =
       sycl::detail::removeDuplicateDevices(Devices);
-  std::shared_ptr<kernel_bundle_impl> sourceImpl = getSyclObjImpl(SourceKB);
+  const std::shared_ptr<kernel_bundle_impl> &sourceImpl =
+      getSyclObjImpl(SourceKB);
   std::shared_ptr<kernel_bundle_impl> KBImpl = sourceImpl->build_from_source(
       UniqueDevices, Options, LogPtr, KernelNames);
   auto result = sycl::detail::createSyclObjFromImpl<exe_kb>(std::move(KBImpl));