[SYCL] Add Level-Zero interop with specification of ownership

sycl/doc/extensions/LevelZeroBackend/LevelZeroBackend.md

@@ -87,7 +87,7 @@ a SYCL object that encapsulates a corresponding Level-Zero object:
 |-------------|:------------|
 |``` make<platform>(ze_driver_handle_t);```|Constructs a SYCL platform instance from a Level-Zero ```ze_driver_handle_t```.|
 |``` make<device>(const platform &, ze_device_handle_t);```|Constructs a SYCL device instance from a Level-Zero ```ze_device_handle_t```. The platform argument gives a SYCL platform, encapsulating a Level-Zero driver supporting the passed Level-Zero device.|
-|``` make<context>(const vector_class<device> &, ze_context_handle_t);```| Constructs a SYCL context instance from a Level-Zero ```ze_context_handle_t```. The context is created against the devices passed in. There must be at least one device given and all the devices must be from the same SYCL platform and thus from the same Level-Zero driver.|
+|``` make<context>(const vector_class<device> &, ze_context_handle_t, ownership = transfer);```| Constructs a SYCL context instance from a Level-Zero ```ze_context_handle_t```. The context is created against the devices passed in. There must be at least one device given and all the devices must be from the same SYCL platform and thus from the same Level-Zero driver. The ```ownership``` argument specifies if the SYCL runtime should take ownership of the passed native handle. The default behavior is to transfer the ownership to the SYCL runtime. See section 4.4 for details.|
 |``` make<queue>(const context &, ze_command_queue_handle_t);```| Constructs a SYCL queue instance from a Level-Zero ```ze_command_queue_handle_t```. The context argument must be a valid SYCL context encapsulating a Level-Zero context. The queue is attached to the first device in the passed SYCL context.|
 |``` make<program>(const context &, ze_module_handle_t);```| Constructs a SYCL program instance from a Level-Zero ```ze_module_handle_t```. The context argument must be a valid SYCL context encapsulating a Level-Zero context. The Level-Zero module must be fully linked (i.e. not require further linking through [```zeModuleDynamicLink```](https://spec.oneapi.com/level-zero/latest/core/api.html?highlight=zemoduledynamiclink#_CPPv419zeModuleDynamicLink8uint32_tP18ze_module_handle_tP28ze_module_build_log_handle_t)), and thus the SYCL program is created in the "linked" state.|
 
@@ -96,23 +96,43 @@ NOTE: We shall consider adding other interoperability as needed, if possible.
 ### 4.4 Level-Zero handles' ownership and thread-safety
 
 The Level-Zero runtime doesn't do reference-counting of its objects, so it is crucial to adhere to these
-practices of how Level-Zero handles are manged.
+practices of how Level-Zero handles are managed. By default, the ownership is transferred to the SYCL runtime, but
+some interoparability API supports overriding this behavior and keep the ownership in the application.
+Use this enumeration for explicit specification of the ownership:
+``` C++
+namespace sycl {
+namespace level_zero {
+
+enum class ownership { transfer, keep };
+
+} // namesace level_zero
+} // namespace sycl
+```
 
-#### 4.4.1 SYCL runtime takes ownership
+#### 4.4.1 SYCL runtime takes ownership (default)
 
 Whenever the application creates a SYCL object from the corresponding Level-Zero handle via one of the ```make<T>()``` functions,
-the SYCL runtime takes ownership of the Level-Zero handle. The application must not use the Level-Zero handle after
-the last host copy of the SYCL object is destroyed (as described in the core SYCL specification under
-"Common reference semantics"), and the application must not destroy the Level-Zero handle itself.
+the SYCL runtime takes ownership of the Level-Zero handle, if no explicit ```ownership::keep``` was specified.
+The application must not use the Level-Zero handle after the last host copy of the SYCL object is destroyed (
+as described in the core SYCL specification under "Common reference semantics"), and the application must not
+destroy the Level-Zero handle itself.
+
+#### 4.4.2 Application keeps ownership (explicit)
+
+If SYCL object is created with an interoperability API explicitly asking to keep the native handle ownership in the application with
+```ownership::keep``` then the SYCL runtime does not take the ownership and will not destroy the Level-Zero handle at the destruction of the SYCL object.
+The application is responsible for destroying the native handle when it no longer needs it, but it must not destroy the
+handle before the last host copy of the SYCL object is destroyed (as described in the core SYCL specification under
+"Common reference semantics").
 
-#### 4.4.2 SYCL runtime assumes ownership
+#### 4.4.3 Obtaining native handle does not change ownership
 
-The application may call the ```get_native<T>()``` member function of a SYCL object to retrieve the underlying Level-Zero handle,
-however, the SYCL runtime continues to retain ownership of this handle. The application must not use this handle after
-the last host copy of the SYCL object is destroyed (as described in the core SYCL specification under
-"Common reference semantics"), and the application must not destroy the Level-Zero handle.
+The application may call the ```get_native<T>()``` member function of a SYCL object to retrieve the underlying Level-Zero handle.
+Doing so does not change the ownership of the the Level-Zero handle.  Therefore, the application may not use this
+handle after the last host copy of the SYCL object is destroyed (as described in the core SYCL specification under
+"Common reference semantics") unless the SYCL object was created by the application with ```ownership::keep```.
 
-#### 4.4.3 Considerations for multi-threaded environment
+#### 4.4.4 Considerations for multi-threaded environment
 
 The Level-Zero API is not thread-safe, refer to <https://spec.oneapi.com/level-zero/latest/core/INTRO.html#multithreading-and-concurrency>.
 Applications must make sure that the Level-Zero handles themselves aren't used simultaneously from different threads.
@@ -123,4 +143,5 @@ the application should not attempt further direct use of those handles.
 |Rev|Date|Author|Changes|
 |-------------|:------------|:------------|:------------|
 |1|2021-01-26|Sergey Maslov|Initial public working draft
+|2|2021-02-22|Sergey Maslov|Introduced explicit ownership for context
 

sycl/include/CL/sycl/backend/level_zero.hpp

@@ -51,12 +51,20 @@ struct interop<backend::level_zero, accessor<DataT, Dimensions, AccessMode,
 
 namespace level_zero {
 
-// Implementation of various "make" functions resides in libsycl.so
+// Since Level-Zero is not doing any reference counting itself, we have to
+// be explicit about the ownership of the native handles used in the
+// interop functions below.
+//
+enum class ownership { transfer, keep };
+
+// Implementation of various "make" functions resides in libsycl.so and thus
+// their interface needs to be backend agnostic.
 __SYCL_EXPORT platform make_platform(pi_native_handle NativeHandle);
 __SYCL_EXPORT device make_device(const platform &Platform,
                                  pi_native_handle NativeHandle);
 __SYCL_EXPORT context make_context(const vector_class<device> &DeviceList,
-                                   pi_native_handle NativeHandle);
+                                   pi_native_handle NativeHandle,
+                                   bool keep_ownership = false);
 __SYCL_EXPORT program make_program(const context &Context,
                                    pi_native_handle NativeHandle);
 __SYCL_EXPORT queue make_queue(const context &Context,
@@ -82,11 +90,17 @@ T make(const platform &Platform,
 ///        created SYCL context. Provided devices and native context handle must
 ///        be associated with the same platform.
 /// \param Interop is a Level Zero native context handle.
+/// \param Ownership (optional) specifies who will assume ownership of the
+///        native context handle. Default is that SYCL RT does, so it destroys
+///        the native handle when the created SYCL object goes out of life.
+///
 template <typename T, typename std::enable_if<
                           std::is_same<T, context>::value>::type * = nullptr>
 T make(const vector_class<device> &DeviceList,
-       typename interop<backend::level_zero, T>::type Interop) {
-  return make_context(DeviceList, detail::pi::cast<pi_native_handle>(Interop));
+       typename interop<backend::level_zero, T>::type Interop,
+       ownership Ownership = ownership::transfer) {
+  return make_context(DeviceList, detail::pi::cast<pi_native_handle>(Interop),
+                      Ownership == ownership::keep);
 }
 
 // Construction of SYCL program.

sycl/include/CL/sycl/detail/pi.h

@@ -34,9 +34,10 @@
 //   pi_device_binary_property_set PropertySetsBegin;
 //   pi_device_binary_property_set PropertySetsEnd;
 // 2. A number of types needed to define pi_device_binary_property_set added.
+// 3. Added new ownership argument to piextContextCreateWithNativeHandle.
 //
-#define _PI_H_VERSION_MAJOR 2
-#define _PI_H_VERSION_MINOR 3
+#define _PI_H_VERSION_MAJOR 3
+#define _PI_H_VERSION_MINOR 4
 
 #define _PI_STRING_HELPER(a) #a
 #define _PI_CONCAT(a, b) _PI_STRING_HELPER(a.b)
@@ -983,6 +984,8 @@ piextContextGetNativeHandle(pi_context context, pi_native_handle *nativeHandle);
 /// \param devices is the list of devices in the context. Parameter is ignored
 ///        if devices can be queried from the context native handle for a
 ///        backend.
+/// \param ownNativeHandle tells if SYCL RT should assume the ownership of
+///        the native handle, if it can.
 /// \param context is the PI context created from the native handle.
 /// \return PI_SUCCESS if successfully created pi_context from the handle.
 ///         PI_OUT_OF_HOST_MEMORY if can't allocate memory for the pi_context
@@ -991,7 +994,7 @@ piextContextGetNativeHandle(pi_context context, pi_native_handle *nativeHandle);
 ///         native handle. PI_UNKNOWN_ERROR in case of another error.
 __SYCL_EXPORT pi_result piextContextCreateWithNativeHandle(
     pi_native_handle nativeHandle, pi_uint32 numDevices,
-    const pi_device *devices, pi_context *context);
+    const pi_device *devices, bool ownNativeHandle, pi_context *context);
 
 //
 // Queue

sycl/plugins/cuda/pi_cuda.cpp

@@ -1698,6 +1698,7 @@ pi_result cuda_piextContextGetNativeHandle(pi_context context,
 pi_result cuda_piextContextCreateWithNativeHandle(pi_native_handle nativeHandle,
                                                   pi_uint32 num_devices,
                                                   const pi_device *devices,
+                                                  bool ownNativeHandle,
                                                   pi_context *context) {
   cl::sycl::detail::pi::die(
       "Creation of PI context from native handle not implemented");

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -1954,7 +1954,7 @@ pi_result piContextCreate(const pi_context_properties *Properties,
   ZE_CALL(zeContextCreate((*Devices)->Platform->ZeDriver, &ContextDesc,
                           &ZeContext));
   try {
-    *RetContext = new _pi_context(ZeContext, NumDevices, Devices);
+    *RetContext = new _pi_context(ZeContext, NumDevices, Devices, true);
     (*RetContext)->initialize();
   } catch (const std::bad_alloc &) {
     return PI_OUT_OF_HOST_MEMORY;
@@ -2013,6 +2013,7 @@ pi_result piextContextGetNativeHandle(pi_context Context,
 pi_result piextContextCreateWithNativeHandle(pi_native_handle NativeHandle,
                                              pi_uint32 NumDevices,
                                              const pi_device *Devices,
+                                             bool OwnNativeHandle,
                                              pi_context *RetContext) {
   PI_ASSERT(NativeHandle, PI_INVALID_VALUE);
   PI_ASSERT(Devices, PI_INVALID_DEVICE);
@@ -2021,7 +2022,7 @@ pi_result piextContextCreateWithNativeHandle(pi_native_handle NativeHandle,
 
   try {
     *RetContext = new _pi_context(pi_cast<ze_context_handle_t>(NativeHandle),
-                                  NumDevices, Devices);
+                                  NumDevices, Devices, OwnNativeHandle);
     (*RetContext)->initialize();
   } catch (const std::bad_alloc &) {
     return PI_OUT_OF_HOST_MEMORY;
@@ -2045,7 +2046,8 @@ pi_result piContextRelease(pi_context Context) {
   PI_ASSERT(Context, PI_INVALID_CONTEXT);
 
   if (--(Context->RefCount) == 0) {
-    auto ZeContext = Context->ZeContext;
+    ze_context_handle_t DestoryZeContext =
+        Context->OwnZeContext ? Context->ZeContext : nullptr;
 
     // Clean up any live memory associated with Context
     pi_result Result = Context->finalize();
@@ -2059,7 +2061,8 @@ pi_result piContextRelease(pi_context Context) {
     // and therefore it must be valid at that point.
     // Technically it should be placed to the destructor of pi_context
     // but this makes API error handling more complex.
-    ZE_CALL(zeContextDestroy(ZeContext));
+    if (DestoryZeContext)
+      ZE_CALL(zeContextDestroy(DestoryZeContext));
 
     return Result;
   }

sycl/plugins/level_zero/pi_level_zero.hpp

@@ -171,8 +171,9 @@ struct _pi_device : _pi_object {
 
 struct _pi_context : _pi_object {
   _pi_context(ze_context_handle_t ZeContext, pi_uint32 NumDevices,
-              const pi_device *Devs)
-      : ZeContext{ZeContext}, Devices{Devs, Devs + NumDevices},
+              const pi_device *Devs, bool OwnZeContext)
+      : ZeContext{ZeContext},
+        OwnZeContext{OwnZeContext}, Devices{Devs, Devs + NumDevices},
         ZeCommandListInit{nullptr}, ZeEventPool{nullptr},
         NumEventsAvailableInEventPool{}, NumEventsLiveInEventPool{} {
     // Create USM allocator context for each pair (device, context).
@@ -201,6 +202,10 @@ struct _pi_context : _pi_object {
   // resources that may be used by multiple devices.
   ze_context_handle_t ZeContext;
 
+  // Indicates if we own the ZeContext or it came from interop that
+  // asked to not transfer the ownership to SYCL RT.
+  bool OwnZeContext;
+
   // Keep the PI devices this PI context was created for.
   std::vector<pi_device> Devices;
 

sycl/plugins/opencl/pi_opencl.cpp

@@ -536,10 +536,12 @@ pi_result piContextCreate(const pi_context_properties *properties,
 pi_result piextContextCreateWithNativeHandle(pi_native_handle nativeHandle,
                                              pi_uint32 num_devices,
                                              const pi_device *devices,
+                                             bool ownNativeHandle,
                                              pi_context *piContext) {
   (void)num_devices;
   (void)devices;
   assert(piContext != nullptr);
+  assert(ownNativeHandle == false);
   *piContext = reinterpret_cast<pi_context>(nativeHandle);
   return PI_SUCCESS;
 }

sycl/source/backend/level_zero.cpp

@@ -48,7 +48,8 @@ __SYCL_EXPORT device make_device(const platform &Platform,
 //----------------------------------------------------------------------------
 // Implementation of level_zero::make<context>
 __SYCL_EXPORT context make_context(const vector_class<device> &DeviceList,
-                                   pi_native_handle NativeHandle) {
+                                   pi_native_handle NativeHandle,
+                                   bool KeepOwnership) {
   const auto &Plugin = pi::getPlugin<backend::level_zero>();
   // Create PI context first.
   pi_context PiContext;
@@ -57,12 +58,19 @@ __SYCL_EXPORT context make_context(const vector_class<device> &DeviceList,
     DeviceHandles.push_back(detail::getSyclObjImpl(Dev)->getHandleRef());
   }
   Plugin.call<PiApiKind::piextContextCreateWithNativeHandle>(
-      NativeHandle, DeviceHandles.size(), DeviceHandles.data(), &PiContext);
+      NativeHandle, DeviceHandles.size(), DeviceHandles.data(), !KeepOwnership,
+      &PiContext);
   // Construct the SYCL context from PI context.
   return detail::createSyclObjFromImpl<context>(
       std::make_shared<context_impl>(PiContext, async_handler{}, Plugin));
 }
 
+// TODO: remove this version (without ownership) when allowed to break ABI.
+__SYCL_EXPORT context make_context(const vector_class<device> &DeviceList,
+                                   pi_native_handle NativeHandle) {
+  return make_context(DeviceList, NativeHandle, false);
+}
+
 //----------------------------------------------------------------------------
 // Implementation of level_zero::make<program>
 __SYCL_EXPORT program make_program(const context &Context,

sycl/source/backend/opencl.cpp

@@ -51,7 +51,7 @@ __SYCL_EXPORT context make_context(pi_native_handle NativeHandle) {
   // Create PI context first.
   pi::PiContext PiContext;
   Plugin.call<PiApiKind::piextContextCreateWithNativeHandle>(
-      NativeHandle, 0, nullptr, &PiContext);
+      NativeHandle, 0, nullptr, false, &PiContext);
   // Construct the SYCL context from PI context.
   return detail::createSyclObjFromImpl<context>(
       std::make_shared<context_impl>(PiContext, async_handler{}, Plugin));

sycl/source/detail/context_impl.cpp

@@ -56,7 +56,8 @@ context_impl::context_impl(const vector_class<cl::sycl::device> Devices,
     getPlugin().call<PiApiKind::piContextCreate>(
         Props, DeviceIds.size(), DeviceIds.data(), nullptr, nullptr, &MContext);
 #else
-    cl::sycl::detail::pi::die("CUDA support was not enabled at compilation time");
+    cl::sycl::detail::pi::die(
+        "CUDA support was not enabled at compilation time");
 #endif
   } else {
     getPlugin().call<PiApiKind::piContextCreate>(nullptr, DeviceIds.size(),
@@ -96,7 +97,12 @@ context_impl::context_impl(RT::PiContext PiContext, async_handler AsyncHandler,
   // TODO catch an exception and put it to list of asynchronous exceptions
   // getPlugin() will be the same as the Plugin passed. This should be taken
   // care of when creating device object.
-  getPlugin().call<PiApiKind::piContextRetain>(MContext);
+  //
+  // TODO: Move this backend-specific retain of the context to SYCL-2020 style
+  //       make_context<backend::opencl> interop, when that is created.
+  if (getPlugin().getBackend() == cl::sycl::backend::opencl) {
+    getPlugin().call<PiApiKind::piContextRetain>(MContext);
+  }
   MKernelProgramCache.setContextPtr(this);
 }
 
@@ -153,8 +159,8 @@ KernelProgramCache &context_impl::getKernelProgramCache() const {
   return MKernelProgramCache;
 }
 
-bool
-context_impl::hasDevice(shared_ptr_class<detail::device_impl> Device) const {
+bool context_impl::hasDevice(
+    shared_ptr_class<detail::device_impl> Device) const {
   for (auto D : MDevices)
     if (getSyclObjImpl(D) == Device)
       return true;