[SYCL] Use per-kernel mutex for all cases when the kernel is cacheable.

Signed-off-by: Alexander Flegontov <[email protected]>

Author: alexanderfle

sycl/source/detail/program_impl.hpp

@@ -318,6 +318,9 @@ class program_impl {
   /// Tells whether a specialization constant has been set for this program.
   bool hasSetSpecConstants() const { return !SpecConstRegistry.empty(); }
 
+  /// \return true if caching is allowed for this program.
+  bool is_cacheable() const { return MProgramAndKernelCachingAllowed; }
+
 private:
   // Deligating Constructor used in Implementation.
   program_impl(ContextImplPtr Context, pi_native_handle InteropProgram,
@@ -368,9 +371,6 @@ class program_impl {
   /// \return a vector of devices managed by the plugin.
   vector_class<RT::PiDevice> get_pi_devices() const;
 
-  /// \return true if caching is allowed for this program.
-  bool is_cacheable() const { return MProgramAndKernelCachingAllowed; }
-
   /// \param Options is a string containing OpenCL C build options.
   /// \return true if caching is allowed for this program and build options.
   static bool is_cacheable_with_options(const string_class &Options) {

sycl/source/detail/scheduler/commands.cpp

@@ -1804,6 +1804,18 @@ cl_int ExecCGCommand::enqueueImp() {
       assert(ExecKernel->MSyclKernel->get_info<info::kernel::context>() ==
              Context);
       Kernel = ExecKernel->MSyclKernel->getHandleRef();
+
+      auto SyclProg = detail::getSyclObjImpl(
+          ExecKernel->MSyclKernel->get_info<info::kernel::program>());
+      if (SyclProg->is_cacheable()) {
+        RT::PiKernel FoundKernel = nullptr;
+        std::tie(FoundKernel, KernelMutex) =
+            detail::ProgramManager::getInstance().getOrCreateKernel(
+                ExecKernel->MOSModuleHandle,
+                ExecKernel->MSyclKernel->get_info<info::kernel::context>(),
+                ExecKernel->MKernelName, SyclProg.get());
+        assert(FoundKernel == Kernel);
+      }
     } else {
       std::tie(Kernel, KernelMutex) =
           detail::ProgramManager::getInstance().getOrCreateKernel(
@@ -1873,6 +1885,7 @@ cl_int ExecCGCommand::enqueueImp() {
 
     pi_result Error = PI_SUCCESS;
     if (KernelMutex != nullptr) {
+      // For cacheable kernels, we use per-kernel mutex
       std::lock_guard<std::mutex> Lock(*KernelMutex);
       Error = SetKernelParamsAndLaunch();
     } else {