Basic no-handler single task support

Author: slawekptak

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>
@@ -262,7 +266,7 @@ void nd_launch(queue Q, nd_range<Dimensions> Range, const KernelType &KernelObj,
 #ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
   // TODO The handler-less path does not support reductions yet.
   if constexpr (sizeof...(ReductionsT) == 0) {
-    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
@@ -298,7 +302,7 @@ void nd_launch(queue Q, launch_config<nd_range<Dimensions>, Properties> Config,
     ext::oneapi::experimental::detail::LaunchConfigAccess<nd_range<Dimensions>,
                                                           Properties>
         ConfigAccess(Config);
-    detail::submit_kernel_direct<KernelName>(
+    detail::submit_kernel_direct_parallel_for<KernelName>(
         std::move(Q), ConfigAccess.getProperties(), ConfigAccess.getRange(),
         KernelObj);
   } else

sycl/include/sycl/khr/free_function_commands.hpp

@@ -158,7 +158,7 @@ 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
-  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
@@ -173,7 +173,7 @@ 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
-  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
@@ -188,7 +188,7 @@ 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
-  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
@@ -305,7 +305,10 @@ 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.
@@ -3278,7 +3332,7 @@ class __SYCL_EXPORT queue : public detail::OwnerLessBase<queue> {
 #ifdef __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT
     // TODO The handler-less path does not support reductions yet.
     if constexpr (sizeof...(RestT) == 1) {
-      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