[NFC][SYCL] Pass queue_impl by raw ptr/ref in misc files (#19006)

Continuation of the refactoring efforts in
#18715
#18748
#18830
#18907
#18983

Author: aelovikov-intel

sycl/source/detail/helpers.hpp

@@ -21,7 +21,6 @@ class event;
 namespace detail {
 class CGExecKernel;
 class queue_impl;
-using QueueImplPtr = std::shared_ptr<sycl::detail::queue_impl>;
 class RTDeviceBinaryImage;
 
 #ifdef __INTEL_PREVIEW_BREAKING_CHANGES

sycl/source/detail/memory_manager.hpp

@@ -28,7 +28,6 @@ class queue_impl;
 class event_impl;
 class context_impl;
 
-using QueueImplPtr = std::shared_ptr<detail::queue_impl>;
 using EventImplPtr = std::shared_ptr<detail::event_impl>;
 
 // The class contains methods that work with memory. All operations with

sycl/source/enqueue_functions.cpp

@@ -15,26 +15,25 @@ namespace ext::oneapi::experimental {
 
 __SYCL_EXPORT void memcpy(queue Q, void *Dest, const void *Src, size_t NumBytes,
                           const sycl::detail::code_location &CodeLoc) {
-  sycl::detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
-  auto QueueImplPtr = sycl::detail::getSyclObjImpl(Q);
-  QueueImplPtr->memcpy(Dest, Src, NumBytes, {},
-                       /*CallerNeedsEvent=*/false, TlsCodeLocCapture.query());
+  detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
+  detail::getSyclObjImpl(Q)->memcpy(Dest, Src, NumBytes, {},
+                                    /*CallerNeedsEvent=*/false,
+                                    TlsCodeLocCapture.query());
 }
 
 __SYCL_EXPORT void memset(queue Q, void *Ptr, int Value, size_t NumBytes,
                           const sycl::detail::code_location &CodeLoc) {
-  sycl::detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
-  auto QueueImplPtr = sycl::detail::getSyclObjImpl(Q);
-  QueueImplPtr->memset(Ptr, Value, NumBytes, {},
-                       /*CallerNeedsEvent=*/false);
+  detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
+  detail::getSyclObjImpl(Q)->memset(Ptr, Value, NumBytes, {},
+                                    /*CallerNeedsEvent=*/false);
 }
 
 __SYCL_EXPORT void mem_advise(queue Q, void *Ptr, size_t NumBytes, int Advice,
                               const sycl::detail::code_location &CodeLoc) {
-  sycl::detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
-  auto QueueImplPtr = sycl::detail::getSyclObjImpl(Q);
-  QueueImplPtr->mem_advise(Ptr, NumBytes, ur_usm_advice_flags_t(Advice), {},
-                           /*CallerNeedsEvent=*/false);
+  detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
+  detail::getSyclObjImpl(Q)->mem_advise(Ptr, NumBytes,
+                                        ur_usm_advice_flags_t(Advice), {},
+                                        /*CallerNeedsEvent=*/false);
 }
 
 } // namespace ext::oneapi::experimental

sycl/source/queue.cpp

@@ -328,15 +328,15 @@ void queue::wait_and_throw_proxy(const detail::code_location &CodeLoc) {
 }
 
 static event
-getBarrierEventForInorderQueueHelper(const detail::QueueImplPtr QueueImpl) {
+getBarrierEventForInorderQueueHelper(detail::queue_impl &QueueImpl) {
   // This function should not be called when a queue is recording to a graph,
   // as a graph can record from multiple queues and we cannot guarantee the
   // last node added by an in-order queue will be the last node added to the
   // graph.
-  assert(!QueueImpl->hasCommandGraph() &&
+  assert(!QueueImpl.hasCommandGraph() &&
          "Should not be called in on graph recording.");
 
-  sycl::detail::optional<event> LastEvent = QueueImpl->getLastEvent();
+  sycl::detail::optional<event> LastEvent = QueueImpl.getLastEvent();
   if (LastEvent)
     return *LastEvent;
 
@@ -353,11 +353,7 @@ getBarrierEventForInorderQueueHelper(const detail::QueueImplPtr QueueImpl) {
 /// \return a SYCL event object, which corresponds to the queue the command
 /// group is being enqueued on.
 event queue::ext_oneapi_submit_barrier(const detail::code_location &CodeLoc) {
-  if (is_in_order() && !impl->hasCommandGraph() && !impl->MIsProfilingEnabled) {
-    return getBarrierEventForInorderQueueHelper(impl);
-  }
-
-  return submit([=](handler &CGH) { CGH.ext_oneapi_barrier(); }, CodeLoc);
+  return ext_oneapi_submit_barrier(std::vector<event>{}, CodeLoc);
 }
 
 /// Prevents any commands submitted afterward to this queue from executing
@@ -379,11 +375,14 @@ event queue::ext_oneapi_submit_barrier(const std::vector<event> &WaitList,
       });
   if (is_in_order() && !impl->hasCommandGraph() && !impl->MIsProfilingEnabled &&
       AllEventsEmptyOrNop) {
-    return getBarrierEventForInorderQueueHelper(impl);
+    return getBarrierEventForInorderQueueHelper(*impl);
   }
 
-  return submit([=](handler &CGH) { CGH.ext_oneapi_barrier(WaitList); },
-                CodeLoc);
+  if (WaitList.empty())
+    return submit([=](handler &CGH) { CGH.ext_oneapi_barrier(); }, CodeLoc);
+  else
+    return submit([=](handler &CGH) { CGH.ext_oneapi_barrier(WaitList); },
+                  CodeLoc);
 }
 
 template <typename Param>

sycl/unittests/Extensions/USMMemcpy2D.cpp

@@ -280,8 +280,7 @@ TEST(USMMemcpy2DTest, USMMemops2DSupported) {
   sycl::platform Plt = sycl::platform();
   sycl::queue Q{Plt.get_devices()[0]};
 
-  std::shared_ptr<sycl::detail::queue_impl> QueueImpl =
-      sycl::detail::getSyclObjImpl(Q);
+  sycl::detail::queue_impl &QueueImpl = *sycl::detail::getSyclObjImpl(Q);
 
   mock::getCallbacks().set_after_callback(
       "urContextGetInfo", &after_urContextGetInfo<true, true, true>);
@@ -297,7 +296,7 @@ TEST(USMMemcpy2DTest, USMMemops2DSupported) {
 
   Q.ext_oneapi_fill2d(Ptr1, 5, 42l, 4, 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_FILL2D_SUPPORT);
-  EXPECT_EQ(LastFill2D.hQueue, (ur_queue_handle_t)QueueImpl->getHandleRef());
+  EXPECT_EQ(LastFill2D.hQueue, (ur_queue_handle_t)QueueImpl.getHandleRef());
   EXPECT_EQ(LastFill2D.pMem, (void *)Ptr1);
   EXPECT_EQ(LastFill2D.pitch, (size_t)5);
   EXPECT_EQ(LastFill2D.patternSize, sizeof(long));
@@ -306,7 +305,7 @@ TEST(USMMemcpy2DTest, USMMemops2DSupported) {
 
   Q.ext_oneapi_memset2d(Ptr1, 5 * sizeof(long), 123, 4 * sizeof(long), 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_FILL2D_SUPPORT);
-  EXPECT_EQ(LastFill2D.hQueue, (ur_queue_handle_t)QueueImpl->getHandleRef());
+  EXPECT_EQ(LastFill2D.hQueue, (ur_queue_handle_t)QueueImpl.getHandleRef());
   EXPECT_EQ(LastFill2D.pMem, (void *)Ptr1);
   EXPECT_EQ(LastFill2D.pitch, (size_t)5 * sizeof(long));
   EXPECT_EQ(LastFill2D.pattern[0], 123);
@@ -316,7 +315,7 @@ TEST(USMMemcpy2DTest, USMMemops2DSupported) {
   Q.ext_oneapi_memcpy2d(Ptr1, 5 * sizeof(long), Ptr2, 8 * sizeof(long),
                         4 * sizeof(long), 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_MEMCPY2D_SUPPORT);
-  EXPECT_EQ(LastMemcpy2D.hQueue, (ur_queue_handle_t)QueueImpl->getHandleRef());
+  EXPECT_EQ(LastMemcpy2D.hQueue, (ur_queue_handle_t)QueueImpl.getHandleRef());
   EXPECT_EQ(LastMemcpy2D.pDst, (void *)Ptr1);
   EXPECT_EQ(LastMemcpy2D.dstPitch, (size_t)5 * sizeof(long));
   EXPECT_EQ(LastMemcpy2D.pSrc, (void *)Ptr2);
@@ -326,7 +325,7 @@ TEST(USMMemcpy2DTest, USMMemops2DSupported) {
 
   Q.ext_oneapi_copy2d(Ptr1, 5, Ptr2, 8, 4, 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_MEMCPY2D_SUPPORT);
-  EXPECT_EQ(LastMemcpy2D.hQueue, (ur_queue_handle_t)QueueImpl->getHandleRef());
+  EXPECT_EQ(LastMemcpy2D.hQueue, (ur_queue_handle_t)QueueImpl.getHandleRef());
   EXPECT_EQ(LastMemcpy2D.pDst, (void *)Ptr2);
   EXPECT_EQ(LastMemcpy2D.dstPitch, (size_t)8 * sizeof(long));
   EXPECT_EQ(LastMemcpy2D.pSrc, (void *)Ptr1);
@@ -381,8 +380,7 @@ TEST(USMMemcpy2DTest, USMFillSupportedOnly) {
   sycl::platform Plt = sycl::platform();
   sycl::queue Q{Plt.get_devices()[0]};
 
-  std::shared_ptr<sycl::detail::queue_impl> QueueImpl =
-      sycl::detail::getSyclObjImpl(Q);
+  sycl::detail::queue_impl &QueueImpl = *sycl::detail::getSyclObjImpl(Q);
 
   mock::getCallbacks().set_after_callback(
       "urContextGetInfo", &after_urContextGetInfo<true, false, false>);
@@ -402,7 +400,7 @@ TEST(USMMemcpy2DTest, USMFillSupportedOnly) {
 
   Q.ext_oneapi_fill2d(Ptr1, 5, 42l, 4, 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_FILL2D_SUPPORT);
-  EXPECT_EQ(LastFill2D.hQueue, QueueImpl->getHandleRef());
+  EXPECT_EQ(LastFill2D.hQueue, QueueImpl.getHandleRef());
   EXPECT_EQ(LastFill2D.pMem, (void *)Ptr1);
   EXPECT_EQ(LastFill2D.pitch, (size_t)5);
   EXPECT_EQ(LastFill2D.patternSize, sizeof(long));
@@ -427,8 +425,7 @@ TEST(USMMemcpy2DTest, USMMemsetSupportedOnly) {
   sycl::platform Plt = sycl::platform();
   sycl::queue Q{Plt.get_devices()[0]};
 
-  std::shared_ptr<sycl::detail::queue_impl> QueueImpl =
-      sycl::detail::getSyclObjImpl(Q);
+  sycl::detail::queue_impl &QueueImpl = *sycl::detail::getSyclObjImpl(Q);
 
   // Enable fill + set, they are implemented with the same entry point in the
   // backend so supporting one means supporting both.
@@ -450,7 +447,7 @@ TEST(USMMemcpy2DTest, USMMemsetSupportedOnly) {
 
   Q.ext_oneapi_memset2d(Ptr1, 5 * sizeof(long), 123, 4 * sizeof(long), 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_FILL2D_SUPPORT);
-  EXPECT_EQ(LastFill2D.hQueue, QueueImpl->getHandleRef());
+  EXPECT_EQ(LastFill2D.hQueue, QueueImpl.getHandleRef());
   EXPECT_EQ(LastFill2D.pMem, (void *)Ptr1);
   EXPECT_EQ(LastFill2D.pitch, (size_t)5 * sizeof(long));
   EXPECT_EQ(LastFill2D.pattern[0], 123);
@@ -475,8 +472,7 @@ TEST(USMMemcpy2DTest, USMMemcpySupportedOnly) {
   sycl::platform Plt = sycl::platform();
   sycl::queue Q{Plt.get_devices()[0]};
 
-  std::shared_ptr<sycl::detail::queue_impl> QueueImpl =
-      sycl::detail::getSyclObjImpl(Q);
+  sycl::detail::queue_impl &QueueImpl = *sycl::detail::getSyclObjImpl(Q);
 
   mock::getCallbacks().set_after_callback(
       "urContextGetInfo", &after_urContextGetInfo<false, false, true>);
@@ -505,7 +501,7 @@ TEST(USMMemcpy2DTest, USMMemcpySupportedOnly) {
   Q.ext_oneapi_memcpy2d(Ptr1, 5 * sizeof(long), Ptr2, 8 * sizeof(long),
                         4 * sizeof(long), 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_MEMCPY2D_SUPPORT);
-  EXPECT_EQ(LastMemcpy2D.hQueue, QueueImpl->getHandleRef());
+  EXPECT_EQ(LastMemcpy2D.hQueue, QueueImpl.getHandleRef());
   EXPECT_EQ(LastMemcpy2D.pDst, (void *)Ptr1);
   EXPECT_EQ(LastMemcpy2D.dstPitch, (size_t)5 * sizeof(long));
   EXPECT_EQ(LastMemcpy2D.pSrc, (void *)Ptr2);
@@ -516,7 +512,7 @@ TEST(USMMemcpy2DTest, USMMemcpySupportedOnly) {
 
   Q.ext_oneapi_copy2d(Ptr1, 5, Ptr2, 8, 4, 2);
   EXPECT_TRUE(LastMemopsQuery == UR_CONTEXT_INFO_USM_MEMCPY2D_SUPPORT);
-  EXPECT_EQ(LastMemcpy2D.hQueue, QueueImpl->getHandleRef());
+  EXPECT_EQ(LastMemcpy2D.hQueue, QueueImpl.getHandleRef());
   EXPECT_EQ(LastMemcpy2D.pDst, (void *)Ptr2);
   EXPECT_EQ(LastMemcpy2D.dstPitch, (size_t)8 * sizeof(long));
   EXPECT_EQ(LastMemcpy2D.pSrc, (void *)Ptr1);

sycl/unittests/program_manager/arg_mask/EliminatedArgMask.cpp

@@ -135,8 +135,8 @@ class MockHandler : public sycl::handler {
 public:
   using sycl::handler::impl;
 
-  MockHandler(std::shared_ptr<sycl::detail::queue_impl> Queue)
-      : sycl::handler(Queue, /*CallerNeedsEvent*/ true) {}
+  MockHandler(sycl::detail::queue_impl &Queue)
+      : sycl::handler(Queue.shared_from_this(), /*CallerNeedsEvent*/ true) {}
 
   std::unique_ptr<sycl::detail::CG> finalize() {
     auto CGH = static_cast<sycl::handler *>(this);
@@ -171,7 +171,7 @@ const sycl::detail::KernelArgMask *getKernelArgMaskFromBundle(
   EXPECT_FALSE(ExecBundle.empty()) << "Expect non-empty exec kernel bundle";
 
   // Emulating processing of command group function
-  MockHandler MockCGH(QueueImpl);
+  MockHandler MockCGH(*QueueImpl);
   MockCGH.use_kernel_bundle(ExecBundle);
   MockCGH.single_task<EAMTestKernel>([] {}); // Actual kernel does not matter