[SYCL] Support sycl::kernel_bundle for multi-device scenario

sycl/cmake/modules/FetchUnifiedRuntime.cmake

@@ -116,14 +116,8 @@ if(SYCL_UR_USE_FETCH_CONTENT)
       CACHE PATH "Path to external '${name}' adapter source dir" FORCE)
   endfunction()
 
-  set(UNIFIED_RUNTIME_REPO "https://github.com/oneapi-src/unified-runtime.git")
-  # commit 532a4ecb72da4876cef61a4ae4d638e27ad609d5
-  # Merge: 22962057 d944ff33
-  # Author: aarongreig <[email protected]>
-  # Date:   Mon Sep 30 10:43:10 2024 +0100
-  #     Merge pull request #1924 from Bensuo/fabio/cmd_buffer_kernel_update
-  #     Add support for command-buffer kernel updates
-  set(UNIFIED_RUNTIME_TAG 532a4ecb72da4876cef61a4ae4d638e27ad609d5)
+  set(UNIFIED_RUNTIME_REPO "https://github.com/againull/unified-runtime")
+  set(UNIFIED_RUNTIME_TAG 44a2f0234f9270aab0c8633c07da81e96fc5e214)
 
   set(UMF_BUILD_EXAMPLES OFF CACHE INTERNAL "EXAMPLES")
   # Due to the use of dependentloadflag and no installer for UMF and hwloc we need

sycl/source/detail/context_impl.cpp

@@ -21,6 +21,7 @@
 #include <sycl/property_list.hpp>
 
 #include <algorithm>
+#include <set>
 
 namespace sycl {
 inline namespace _V1 {
@@ -490,7 +491,7 @@ std::optional<ur_program_handle_t> context_impl::getProgramForDevImgs(
     auto &Cache = LockedCache.get().Cache;
     ur_device_handle_t &DevHandle = getSyclObjImpl(Device)->getHandleRef();
     for (std::uintptr_t ImageIDs : ImgIdentifiers) {
-      auto OuterKey = std::make_pair(ImageIDs, DevHandle);
+      auto OuterKey = std::make_pair(ImageIDs, std::set{DevHandle});
       size_t NProgs = KeyMap.count(OuterKey);
       if (NProgs == 0)
         continue;

sycl/source/detail/helpers.cpp

@@ -65,8 +65,8 @@ retrieveKernelBinary(const QueueImplPtr &Queue, const char *KernelName,
     auto DeviceImpl = Queue->getDeviceImplPtr();
     auto Device = detail::createSyclObjFromImpl<device>(DeviceImpl);
     ur_program_handle_t Program =
-        detail::ProgramManager::getInstance().createURProgram(**DeviceImage,
-                                                              Context, Device);
+        detail::ProgramManager::getInstance().createURProgram(
+            **DeviceImage, Context, {Device});
     return {*DeviceImage, Program};
   }
 
@@ -94,7 +94,7 @@ retrieveKernelBinary(const QueueImplPtr &Queue, const char *KernelName,
     DeviceImage = &detail::ProgramManager::getInstance().getDeviceImage(
         KernelName, Context, Device);
     Program = detail::ProgramManager::getInstance().createURProgram(
-        *DeviceImage, Context, Device);
+        *DeviceImage, Context, {Device});
   }
   return {DeviceImage, Program};
 }

sycl/source/detail/kernel_program_cache.hpp

@@ -20,6 +20,7 @@
 #include <atomic>
 #include <condition_variable>
 #include <mutex>
+#include <set>
 #include <type_traits>
 
 #include <boost/unordered/unordered_flat_map.hpp>
@@ -118,9 +119,10 @@ class KernelProgramCache {
    * when debugging environment variables are set and we can just ignore them
    * since all kernels will have their build options overridden with the same
    * string*/
-  using ProgramCacheKeyT =
-      std::pair<std::pair<SerializedObj, std::uintptr_t>, ur_device_handle_t>;
-  using CommonProgramKeyT = std::pair<std::uintptr_t, ur_device_handle_t>;
+  using ProgramCacheKeyT = std::pair<std::pair<SerializedObj, std::uintptr_t>,
+                                     std::set<ur_device_handle_t>>;
+  using CommonProgramKeyT =
+      std::pair<std::uintptr_t, std::set<ur_device_handle_t>>;
 
   struct ProgramCache {
     ::boost::unordered_map<ProgramCacheKeyT, ProgramBuildResultPtr> Cache;