Revert "Improved build results caching for GPU devices."

This reverts commit a1e483a.

Author: bader

sycl/source/detail/program_manager/program_manager.cpp

@@ -476,21 +476,22 @@ RT::PiProgram ProgramManager::getBuiltPIProgram(
   // (e.g. clGetKernelWorkGroupInfo returns CL_INVALID_KERNEL if kernel was
   // created from the program built for sub-device and re-used either on root or
   // other sub-device).
-  // To work around this case we optimize only one case if device type is CPU:
-  // root device shares the same context with its sub-device(s). We built for
-  // the root device and cache the results.
-  // The solution for other devices is to build for any sub-device and use root
-  // device handle as cache key to share build results for any other sub-device
-  // or even a root device.
+  // To work around this case we optimize only one case: root device shares the
+  // same context with its sub-device(s). We built for the root device and
+  // cache the results.
+  // The expected solution is to build for any sub-device and use root device
+  // handle as cache key to share build results for any other sub-device or even
+  // a root device.
+  // TODO: it might be worth testing if Level Zero plug-in supports all cases
+  // and enable more cases for Level Zero.
   DeviceImplPtr Dev = DeviceImpl;
-  if (!Dev->is_gpu())
-    while (!Dev->isRootDevice()) {
-      auto ParentDev =
-          detail::getSyclObjImpl(Dev->get_info<info::device::parent_device>());
-      if (!ContextImpl->hasDevice(ParentDev))
-        break;
-      Dev = ParentDev;
-    }
+  while (!Dev->isRootDevice()) {
+    auto ParentDev =
+        detail::getSyclObjImpl(Dev->get_info<info::device::parent_device>());
+    if (!ContextImpl->hasDevice(ParentDev))
+      break;
+    Dev = ParentDev;
+  }
 
   auto BuildF = [this, &M, &KSId, &ContextImpl, &Dev, Prg, &CompileOpts,
                  &LinkOpts, &JITCompilationIsRequired, SpecConsts] {
@@ -545,16 +546,6 @@ RT::PiProgram ProgramManager::getBuiltPIProgram(
     return BuiltProgram.release();
   };
 
-  // Use root device as a cache key
-  // FIXME: on CPU we can't re-use results unless "Dev" is a root device already
-  // due to Intel OpenCL CPU bug(?). This solution is tested only on Intel GPU
-  // implementation.
-  if (Dev->is_gpu())
-    while (!Dev->isRootDevice()) {
-      auto ParentDev =
-          detail::getSyclObjImpl(Dev->get_info<info::device::parent_device>());
-      Dev = ParentDev;
-    }
   const RT::PiDevice PiDevice = Dev->getHandleRef();
 
   auto BuildResult = getOrBuild<PiProgramT, compile_program_error>(