Temp no handler integration v2

sycl/include/sycl/ext/oneapi/experimental/enqueue_functions.hpp

@@ -152,9 +152,13 @@ template <typename KernelName = sycl::detail::auto_name, typename KernelType>
 void single_task(queue Q, const KernelType &KernelObj,
                  const sycl::detail::code_location &CodeLoc =
                      sycl::detail::code_location::current()) {
+  /*
   submit(
       std::move(Q),
       [&](handler &CGH) { single_task<KernelName>(CGH, KernelObj); }, CodeLoc);
+  */
+  detail::submit_kernel_direct_single_task<KernelName>(std::move(Q), empty_properties_t{},
+                                             KernelObj);
 }
 
 template <typename... ArgsT>
@@ -259,18 +263,15 @@ template <typename KernelName = sycl::detail::auto_name, int Dimensions,
           typename KernelType, typename... ReductionsT>
 void nd_launch(queue Q, nd_range<Dimensions> Range, const KernelType &KernelObj,
                ReductionsT &&...Reductions) {
-#ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
   // TODO The handler-less path does not support reductions and kernel function
   // properties yet.
   if constexpr (sizeof...(ReductionsT) == 0 &&
                 !(ext::oneapi::experimental::detail::
                       HasKernelPropertiesGetMethod<
                           const KernelType &>::value)) {
-    detail::submit_kernel_direct<KernelName>(std::move(Q), empty_properties_t{},
+    detail::submit_kernel_direct_parallel_for<KernelName>(std::move(Q), empty_properties_t{},
                                              Range, KernelObj);
-  } else
-#endif
-  {
+  } else {
     submit(std::move(Q), [&](handler &CGH) {
       nd_launch<KernelName>(CGH, Range, KernelObj,
                             std::forward<ReductionsT>(Reductions)...);

sycl/include/sycl/khr/free_function_commands.hpp

@@ -157,17 +157,14 @@ template <typename KernelType, typename = typename std::enable_if_t<
 void launch_grouped(const queue &q, range<1> r, range<1> size, KernelType &&k,
                     const sycl::detail::code_location &codeLoc =
                         sycl::detail::code_location::current()) {
-#ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
   // TODO The handler-less path does not support kernel function properties yet.
   if constexpr (!(ext::oneapi::experimental::detail::
                       HasKernelPropertiesGetMethod<
                           const KernelType &>::value)) {
-    detail::submit_kernel_direct(
+    detail::submit_kernel_direct_parallel_for(
         q, ext::oneapi::experimental::empty_properties_t{},
         nd_range<1>(r, size), std::forward<KernelType>(k));
-  } else
-#endif
-  {
+  } else {
     submit(
         q, [&](handler &h) { launch_grouped<KernelType>(h, r, size, k); },
         codeLoc);
@@ -178,17 +175,14 @@ template <typename KernelType, typename = typename std::enable_if_t<
 void launch_grouped(const queue &q, range<2> r, range<2> size, KernelType &&k,
                     const sycl::detail::code_location &codeLoc =
                         sycl::detail::code_location::current()) {
-#ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
   // TODO The handler-less path does not support kernel function properties yet.
   if constexpr (!(ext::oneapi::experimental::detail::
                       HasKernelPropertiesGetMethod<
                           const KernelType &>::value)) {
-    detail::submit_kernel_direct(
+    detail::submit_kernel_direct_parallel_for(
         q, ext::oneapi::experimental::empty_properties_t{},
         nd_range<2>(r, size), std::forward<KernelType>(k));
-  } else
-#endif
-  {
+  } else {
     submit(
         q, [&](handler &h) { launch_grouped<KernelType>(h, r, size, k); },
         codeLoc);
@@ -199,17 +193,14 @@ template <typename KernelType, typename = typename std::enable_if_t<
 void launch_grouped(const queue &q, range<3> r, range<3> size, KernelType &&k,
                     const sycl::detail::code_location &codeLoc =
                         sycl::detail::code_location::current()) {
-#ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
   // TODO The handler-less path does not support kernel function properties yet.
   if constexpr (!(ext::oneapi::experimental::detail::
                       HasKernelPropertiesGetMethod<
                           const KernelType &>::value)) {
-    detail::submit_kernel_direct(
+    detail::submit_kernel_direct_parallel_for(
         q, ext::oneapi::experimental::empty_properties_t{},
         nd_range<3>(r, size), std::forward<KernelType>(k));
-  } else
-#endif
-  {
+  } else {
     submit(
         q, [&](handler &h) { launch_grouped<KernelType>(h, r, size, k); },
         codeLoc);
@@ -323,7 +314,10 @@ template <typename KernelType>
 void launch_task(const sycl::queue &q, const KernelType &k,
                  const sycl::detail::code_location &codeLoc =
                      sycl::detail::code_location::current()) {
-  submit(q, [&](handler &h) { launch_task<KernelType>(h, k); }, codeLoc);
+  //submit(q, [&](handler &h) { launch_task<KernelType>(h, k); }, codeLoc);
+  detail::submit_kernel_direct_single_task(q,
+    ext::oneapi::experimental::empty_properties_t{},
+    k, codeLoc);
 }
 
 template <typename... Args>

sycl/include/sycl/queue.hpp

@@ -158,7 +158,7 @@ class __SYCL_EXPORT SubmissionInfo {
 
 template <typename KernelName = detail::auto_name, bool EventNeeded = false,
           typename PropertiesT, typename KernelTypeUniversalRef, int Dims>
-auto submit_kernel_direct(
+auto submit_kernel_direct_parallel_for(
     const queue &Queue, PropertiesT Props, const nd_range<Dims> &Range,
     KernelTypeUniversalRef &&KernelFunc,
     const detail::code_location &CodeLoc = detail::code_location::current()) {
@@ -211,6 +211,53 @@ auto submit_kernel_direct(
   }
 }
 
+template <typename KernelName = detail::auto_name, bool EventNeeded = false,
+          typename PropertiesT, typename KernelTypeUniversalRef>
+auto submit_kernel_direct_single_task(
+    const queue &Queue, PropertiesT Props, KernelTypeUniversalRef &&KernelFunc,
+    const detail::code_location &CodeLoc = detail::code_location::current()) {
+  // TODO Properties not supported yet
+  (void)Props;
+  static_assert(
+      std::is_same_v<PropertiesT,
+                     ext::oneapi::experimental::empty_properties_t>,
+      "Setting properties not supported yet for no-CGH kernel submit.");
+  detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
+
+  using KernelType =
+      std::remove_const_t<std::remove_reference_t<KernelTypeUniversalRef>>;
+
+  using NameT =
+      typename detail::get_kernel_name_t<KernelName, KernelType>::name;
+
+  detail::KernelWrapper<detail::WrapAs::single_task, NameT, KernelType,
+                        void, PropertiesT>::wrap(KernelFunc);
+
+  HostKernelRef<KernelType, KernelTypeUniversalRef, void, 1>
+      HostKernel(std::forward<KernelTypeUniversalRef>(KernelFunc));
+
+  // Instantiating the kernel on the host improves debugging.
+  // Passing this pointer to another translation unit prevents optimization.
+#ifndef NDEBUG
+  // TODO: call library to prevent dropping call due to optimization
+  (void)
+      detail::GetInstantiateKernelOnHostPtr<KernelType, void, 1>();
+#endif
+
+  detail::DeviceKernelInfo *DeviceKernelInfoPtr =
+      &detail::getDeviceKernelInfo<NameT>();
+
+  if constexpr (EventNeeded) {
+    return submit_kernel_direct_with_event_impl(
+        Queue, nd_range<1>{1, 1}, HostKernel, DeviceKernelInfoPtr,
+        TlsCodeLocCapture.query(), TlsCodeLocCapture.isToplevel());
+  } else {
+    submit_kernel_direct_without_event_impl(
+        Queue, nd_range<1>{1, 1}, HostKernel, DeviceKernelInfoPtr,
+        TlsCodeLocCapture.query(), TlsCodeLocCapture.isToplevel());
+  }
+}
+
 } // namespace detail
 
 namespace ext ::oneapi ::experimental {
@@ -2720,14 +2767,21 @@ class __SYCL_EXPORT queue : public detail::OwnerLessBase<queue> {
                                     void(kernel_handler)>::value),
         "sycl::queue.single_task() requires a kernel instead of command group. "
         "Use queue.submit() instead");
-
+/*
     detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
     return submit(
         [&](handler &CGH) {
           CGH.template single_task<KernelName, KernelType, PropertiesT>(
               Properties, KernelFunc);
         },
         TlsCodeLocCapture.query());
+*/
+
+      (void)Properties;
+      return detail::submit_kernel_direct_single_task<KernelName, true>(
+          *this, ext::oneapi::experimental::empty_properties_t{},
+          KernelFunc, CodeLoc);
+
   }
 
   /// single_task version with a kernel represented as a lambda.
@@ -3275,7 +3329,6 @@ class __SYCL_EXPORT queue : public detail::OwnerLessBase<queue> {
   parallel_for(nd_range<Dims> Range, RestT &&...Rest) {
     constexpr detail::code_location CodeLoc = getCodeLocation<KernelName>();
     detail::tls_code_loc_t TlsCodeLocCapture(CodeLoc);
-#ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
     using KernelType = std::tuple_element_t<0, std::tuple<RestT...>>;
 
     // TODO The handler-less path does not support reductions and kernel
@@ -3284,12 +3337,10 @@ class __SYCL_EXPORT queue : public detail::OwnerLessBase<queue> {
                   !(ext::oneapi::experimental::detail::
                         HasKernelPropertiesGetMethod<
                             const KernelType &>::value)) {
-      return detail::submit_kernel_direct<KernelName, true>(
+      return detail::submit_kernel_direct_parallel_for<KernelName, true>(
           *this, ext::oneapi::experimental::empty_properties_t{}, Range,
           Rest...);
-    } else
-#endif
-    {
+    } else {
       return submit(
           [&](handler &CGH) {
             CGH.template parallel_for<KernelName>(Range, Rest...);

sycl/source/detail/queue_impl.cpp

@@ -662,6 +662,14 @@ queue_impl::submit_direct(bool CallerNeedsEvent,
            : true) &&
       !hasCommandGraph();
 
+  if (isInOrder()) {
+    if (SchedulerBypass) {
+      MNoLastEventMode.store(true, std::memory_order_relaxed);
+    } else {
+      MNoLastEventMode.store(false, std::memory_order_relaxed);
+    }
+  }
+
   EventImplPtr EventImpl = SubmitCommandFunc(CGData, SchedulerBypass);
 
   // Sync with the last event for in order queue. For scheduler-bypass flow,

sycl/test-e2e/Basic/test_num_kernel_copies.cpp

@@ -29,7 +29,7 @@ int main(int argc, char **argv) {
 
   kernel<1> krn1;
   q.parallel_for(sycl::nd_range<1>{1, 1}, krn1);
-  assert(copy_count == 1);
+  assert(copy_count == 0);
   assert(move_count == 0);
   copy_count = 0;
 

sycl/unittests/Extensions/FreeFunctionCommands/FreeFunctionCommandsEvents.cpp

@@ -227,7 +227,6 @@ TEST_F(FreeFunctionCommandsEventsTests, LaunchGroupedShortcutNoEvent) {
   ASSERT_EQ(counter_urEnqueueKernelLaunch, size_t{1});
 }
 
-#if __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
 TEST_F(FreeFunctionCommandsEventsTests,
        LaunchGroupedShortcutMoveKernelNoEvent) {
   mock::getCallbacks().set_replace_callback("urEnqueueKernelLaunch",
@@ -276,7 +275,6 @@ TEST_F(FreeFunctionCommandsEventsTests,
   ASSERT_EQ(TestMoveFunctor::MoveCtorCalls, 1);
   ASSERT_EQ(counter_urEnqueueKernelLaunch, size_t{2});
 }
-#endif
 
 TEST_F(FreeFunctionCommandsEventsTests, SubmitLaunchGroupedKernelNoEvent) {
   mock::getCallbacks().set_replace_callback("urEnqueueKernelLaunch",

sycl/unittests/Extensions/FreeFunctionCommands/FreeFunctionEventsHelpers.hpp

@@ -26,53 +26,53 @@ inline ur_result_t after_urKernelGetInfo(void *pParams) {
 static thread_local size_t counter_urEnqueueKernelLaunch = 0;
 inline ur_result_t redefined_urEnqueueKernelLaunch(void *pParams) {
   ++counter_urEnqueueKernelLaunch;
-// TODO The no-handler scheduler submission includes a fix for the event return,
-// where the event is returned by the scheduler on every submission. This fix
-// is not yet applied to the handler-based path.
-#ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
-  auto params = *static_cast<ur_enqueue_kernel_launch_params_t *>(pParams);
-  EXPECT_EQ(*params.pphEvent, nullptr);
-#endif
+  // TODO The no-handler scheduler submission includes a fix for the event
+  // return, where the event is returned by the scheduler on every submission.
+  // This fix is not yet applied to the handler-based path. #ifndef
+  // __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
+  //   auto params = *static_cast<ur_enqueue_kernel_launch_params_t *>(pParams);
+  //   EXPECT_EQ(*params.pphEvent, nullptr);
+  // #endif
   return UR_RESULT_SUCCESS;
 }
 
 static thread_local size_t counter_urUSMEnqueueMemcpy = 0;
 inline ur_result_t redefined_urUSMEnqueueMemcpy(void *pParams) {
   ++counter_urUSMEnqueueMemcpy;
-#ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
-  auto params = *static_cast<ur_enqueue_usm_memcpy_params_t *>(pParams);
-  EXPECT_EQ(*params.pphEvent, nullptr);
-#endif
+  // #ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
+  //   auto params = *static_cast<ur_enqueue_usm_memcpy_params_t *>(pParams);
+  //   EXPECT_EQ(*params.pphEvent, nullptr);
+  // #endif
   return UR_RESULT_SUCCESS;
 }
 
 static thread_local size_t counter_urUSMEnqueueFill = 0;
 inline ur_result_t redefined_urUSMEnqueueFill(void *pParams) {
   ++counter_urUSMEnqueueFill;
-#ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
-  auto params = *static_cast<ur_enqueue_usm_fill_params_t *>(pParams);
-  EXPECT_EQ(*params.pphEvent, nullptr);
-#endif
+  // #ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
+  //   auto params = *static_cast<ur_enqueue_usm_fill_params_t *>(pParams);
+  //   EXPECT_EQ(*params.pphEvent, nullptr);
+  // #endif
   return UR_RESULT_SUCCESS;
 }
 
 static thread_local size_t counter_urUSMEnqueuePrefetch = 0;
 inline ur_result_t redefined_urUSMEnqueuePrefetch(void *pParams) {
   ++counter_urUSMEnqueuePrefetch;
-#ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
-  auto params = *static_cast<ur_enqueue_usm_prefetch_params_t *>(pParams);
-  EXPECT_EQ(*params.pphEvent, nullptr);
-#endif
+  // #ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
+  //   auto params = *static_cast<ur_enqueue_usm_prefetch_params_t *>(pParams);
+  //   EXPECT_EQ(*params.pphEvent, nullptr);
+  // #endif
   return UR_RESULT_SUCCESS;
 }
 
 static thread_local size_t counter_urUSMEnqueueMemAdvise = 0;
 inline ur_result_t redefined_urUSMEnqueueMemAdvise(void *pParams) {
   ++counter_urUSMEnqueueMemAdvise;
-#ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
-  auto params = *static_cast<ur_enqueue_usm_advise_params_t *>(pParams);
-  EXPECT_EQ(*params.pphEvent, nullptr);
-#endif
+  // #ifndef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
+  //   auto params = *static_cast<ur_enqueue_usm_advise_params_t *>(pParams);
+  //   EXPECT_EQ(*params.pphEvent, nullptr);
+  // #endif
   return UR_RESULT_SUCCESS;
 }