[NFC][SYCL] Refactor code around device_image_impl::KernelIDs (#19516)

* Clarify the reason for `std::shared_ptr` (performance, 
  see #5496)
* Clarify that `get_kernel_ids` is preferred over `get_kernel_ids_ptr`
and
  update some of the existing uses of the latter to use the former
* Change `get_kernel_ids` to return `iterator_range` instead of the
  reference to the underlying container. That way we can also return an
  empty range when `shared_ptr` isn't initialized, which was possible
before but many ctors led to believe that is guaranteed to never happen.
* Change those ctors to keep empty `shared_ptr` instead of creating
  empty `vector`s

Author: aelovikov-intel

sycl/source/detail/device_image_impl.hpp

@@ -311,9 +311,7 @@ class device_image_impl
                     private_tag)
       : MBinImage(BinImage), MContext(std::move(Context)),
         MDevices(Devices.to<std::vector<device_impl *>>()), MState(State),
-        MProgram(Program),
-        MKernelIDs(std::make_shared<std::vector<kernel_id>>()),
-        MKernelNames{std::move(KernelNames)},
+        MProgram(Program), MKernelNames{std::move(KernelNames)},
         MEliminatedKernelArgMasks{std::move(EliminatedKernelArgMasks)},
         MSpecConstsDefValBlob(getSpecConstsDefValBlob()),
         MOrigins(ImageOriginKernelCompiler),
@@ -347,7 +345,6 @@ class device_image_impl
       : MBinImage(Src), MContext(std::move(Context)),
         MDevices(Devices.to<std::vector<device_impl *>>()),
         MState(bundle_state::ext_oneapi_source), MProgram(nullptr),
-        MKernelIDs(std::make_shared<std::vector<kernel_id>>()),
         MSpecConstsDefValBlob(getSpecConstsDefValBlob()),
         MOrigins(ImageOriginKernelCompiler),
         MRTCBinInfo(
@@ -361,7 +358,6 @@ class device_image_impl
       : MBinImage(Bytes), MContext(std::move(Context)),
         MDevices(Devices.to<std::vector<device_impl *>>()),
         MState(bundle_state::ext_oneapi_source), MProgram(nullptr),
-        MKernelIDs(std::make_shared<std::vector<kernel_id>>()),
         MSpecConstsDefValBlob(getSpecConstsDefValBlob()),
         MOrigins(ImageOriginKernelCompiler),
         MRTCBinInfo(KernelCompilerBinaryInfo{Lang}) {
@@ -375,9 +371,7 @@ class device_image_impl
       : MBinImage(static_cast<const RTDeviceBinaryImage *>(nullptr)),
         MContext(std::move(Context)),
         MDevices(Devices.to<std::vector<device_impl *>>()), MState(State),
-        MProgram(Program),
-        MKernelIDs(std::make_shared<std::vector<kernel_id>>()),
-        MKernelNames{std::move(KernelNames)},
+        MProgram(Program), MKernelNames{std::move(KernelNames)},
         MSpecConstsDefValBlob(getSpecConstsDefValBlob()),
         MOrigins(ImageOriginKernelCompiler),
         MRTCBinInfo(KernelCompilerBinaryInfo{Lang}) {}
@@ -389,6 +383,8 @@ class device_image_impl
   }
 
   bool has_kernel(const kernel_id &KernelIDCand) const noexcept {
+    if (!MKernelIDs)
+      return false;
     return std::binary_search(MKernelIDs->begin(), MKernelIDs->end(),
                               KernelIDCand, LessByHash<kernel_id>{});
   }
@@ -414,8 +410,18 @@ class device_image_impl
     return false;
   }
 
-  const std::vector<kernel_id> &get_kernel_ids() const noexcept {
-    return *MKernelIDs;
+  iterator_range<std::vector<kernel_id>::const_iterator>
+  get_kernel_ids() const noexcept {
+    if (MKernelIDs)
+      return *MKernelIDs;
+    else
+      return {};
+  }
+  // This should only be used when creating new device_image_impls that have the
+  // exact same set of kernels as the source one. In all other scenarios the
+  // getter above is the one needed:
+  std::shared_ptr<std::vector<kernel_id>> &get_kernel_ids_ptr() noexcept {
+    return MKernelIDs;
   }
 
   bool has_specialization_constants() const noexcept {
@@ -563,10 +569,6 @@ class device_image_impl
 
   const context &get_context() const noexcept { return MContext; }
 
-  std::shared_ptr<std::vector<kernel_id>> &get_kernel_ids_ptr() noexcept {
-    return MKernelIDs;
-  }
-
   std::vector<unsigned char> &get_spec_const_blob_ref() noexcept {
     return MSpecConstsBlob;
   }
@@ -1300,7 +1302,9 @@ class device_image_impl
   ur_program_handle_t MProgram = nullptr;
 
   // List of kernel ids available in this image, elements should be sorted
-  // according to LessByNameComp
+  // according to LessByNameComp. Shared between images for performance reasons
+  // (e.g. when we compile a single image it keeps the same kernels in it as the
+  // original source image).
   std::shared_ptr<std::vector<kernel_id>> MKernelIDs;
 
   // List of known kernel names.

sycl/source/detail/helpers.hpp

@@ -51,6 +51,7 @@ class variadic_iterator {
   using pointer = value_type *;
   static_assert(std::is_same_v<reference, value_type &>);
 
+  variadic_iterator() = default;
   variadic_iterator(const variadic_iterator &) = default;
   variadic_iterator(variadic_iterator &&) = default;
   variadic_iterator(variadic_iterator &) = default;
@@ -88,7 +89,6 @@ class variadic_iterator {
 // Non-owning!
 template <typename iterator> class iterator_range {
   using value_type = typename iterator::value_type;
-  using sycl_type = typename iterator::sycl_type;
 
   template <typename Container, typename = void>
   struct has_reserve : public std::false_type {};
@@ -104,16 +104,20 @@ template <typename iterator> class iterator_range {
   iterator_range(IterTy Begin, IterTy End, size_t Size)
       : Begin(Begin), End(End), Size(Size) {}
 
-  iterator_range()
-      : iterator_range(static_cast<value_type *>(nullptr),
-                       static_cast<value_type *>(nullptr), 0) {}
+  iterator_range() : iterator_range(iterator{}, iterator{}, 0) {}
 
-  template <typename ContainerTy>
+  template <typename ContainerTy, typename = std::void_t<decltype(iterator{
+                                      std::declval<ContainerTy>().begin()})>>
   iterator_range(const ContainerTy &Container)
       : iterator_range(Container.begin(), Container.end(), Container.size()) {}
 
   iterator_range(value_type &Obj) : iterator_range(&Obj, &Obj + 1, 1) {}
 
+  template <typename sycl_type,
+            typename = std::void_t<decltype(iterator{
+                &*getSyclObjImpl(std::declval<sycl_type>())})>,
+            // To make it different from `ContainerTy` overload above:
+            typename = void>
   iterator_range(const sycl_type &Obj)
       : iterator_range(&*getSyclObjImpl(Obj), (&*getSyclObjImpl(Obj) + 1), 1) {}
 
@@ -123,13 +127,15 @@ template <typename iterator> class iterator_range {
   bool empty() const { return Size == 0; }
   decltype(auto) front() const { return *begin(); }
 
-  template <typename Container>
-  std::enable_if_t<
-      check_type_in_v<Container, std::vector<sycl_type>,
-                      std::queue<value_type *>, std::vector<value_type *>,
-                      std::vector<std::shared_ptr<value_type>>>,
-      Container>
-  to() const {
+  // Only enable for ranges of `variadic_iterator` and for the containers with
+  // proper `value_type`. The last part is important so that descendent
+  // `devices_range` could provide its own specialization for
+  // `to<std::vector<device_handle_t>>()`.
+  template <typename Container, typename iterator_ = iterator,
+            typename = std::enable_if_t<check_type_in_v<
+                typename Container::value_type, value_type *,
+                std::shared_ptr<value_type>, typename iterator_::sycl_type>>>
+  Container to() const {
     std::conditional_t<std::is_same_v<Container, std::queue<value_type *>>,
                        typename std::queue<value_type *>::container_type,
                        Container>
@@ -138,31 +144,30 @@ template <typename iterator> class iterator_range {
       Result.reserve(size());
     std::transform(
         begin(), end(), std::back_inserter(Result), [](value_type &E) {
-          if constexpr (std::is_same_v<Container, std::vector<sycl_type>>)
-            return createSyclObjFromImpl<sycl_type>(E);
-          else if constexpr (std::is_same_v<
-                                 Container,
-                                 std::vector<std::shared_ptr<value_type>>>)
+          using container_value_type = typename Container::value_type;
+          if constexpr (std::is_same_v<container_value_type,
+                                       std::shared_ptr<value_type>>)
             return E.shared_from_this();
-          else
+          else if constexpr (std::is_same_v<container_value_type, value_type *>)
             return &E;
+          else
+            return createSyclObjFromImpl<container_value_type>(E);
         });
     if constexpr (std::is_same_v<Container, decltype(Result)>)
       return Result;
     else
       return Container{std::move(Result)};
   }
 
+  // Only enable for ranges of `variadic_iterator` above.
+  template <typename T = iterator,
+            typename = std::void_t<typename T::sycl_type>>
   bool contains(value_type &Other) const {
     return std::find_if(begin(), end(), [&Other](value_type &Elem) {
              return &Elem == &Other;
            }) != end();
   }
 
-protected:
-  template <typename Container>
-  static constexpr bool has_reserve_v = has_reserve<Container>::value;
-
 private:
   iterator Begin;
   iterator End;

sycl/source/detail/kernel_bundle_impl.hpp

@@ -782,7 +782,7 @@ class kernel_bundle_impl
       if (DevImgImpl.getRTCInfo())
         continue;
 
-      const std::vector<kernel_id> &KernelIDs = DevImgImpl.get_kernel_ids();
+      auto KernelIDs = DevImgImpl.get_kernel_ids();
 
       Result.insert(Result.end(), KernelIDs.begin(), KernelIDs.end());
     }

sycl/source/detail/program_manager/program_manager.cpp

@@ -2683,9 +2683,9 @@ ProgramManager::createDependencyImage(const context &Ctx, devices_range Devs,
 
   assert(DepState == getBinImageState(DepImage) &&
          "State mismatch between main image and its dependency");
-  DeviceImageImplPtr DepImpl =
-      device_image_impl::create(DepImage, Ctx, Devs, DepState, DepKernelIDs,
-                                /*PIProgram=*/nullptr, ImageOriginSYCLOffline);
+  DeviceImageImplPtr DepImpl = device_image_impl::create(
+      DepImage, Ctx, Devs, DepState, std::move(DepKernelIDs),
+      /*PIProgram=*/nullptr, ImageOriginSYCLOffline);
 
   return createSyclObjFromImpl<device_image_plain>(std::move(DepImpl));
 }
@@ -2898,10 +2898,8 @@ mergeImageData(const std::vector<device_image_plain> &Imgs,
   for (const device_image_plain &Img : Imgs) {
     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(),
-                       DeviceImageImpl.get_kernel_ids_ptr()->begin(),
-                       DeviceImageImpl.get_kernel_ids_ptr()->end());
+    KernelIDs.insert(KernelIDs.end(), DeviceImageImpl.get_kernel_ids().begin(),
+                     DeviceImageImpl.get_kernel_ids().end());
     // To be able to answer queries about specialziation constants, the new
     // device image should have the specialization constants from all the linked
     // images.

sycl/source/kernel_bundle.cpp

@@ -288,8 +288,8 @@ bool has_kernel_bundle_impl(const context &Ctx, const std::vector<device> &Devs,
       const std::shared_ptr<device_image_impl> &DeviceImageImpl =
           getSyclObjImpl(DeviceImage);
 
-      CombinedKernelIDs.insert(DeviceImageImpl->get_kernel_ids_ptr()->begin(),
-                               DeviceImageImpl->get_kernel_ids_ptr()->end());
+      CombinedKernelIDs.insert(DeviceImageImpl->get_kernel_ids().begin(),
+                               DeviceImageImpl->get_kernel_ids().end());
     }
   }