[SYCL] Update graph constructor/finalize to current spec (#140)

- Add device and context params to graph constructor
- Remove context from finalize
- Minor changes to graph_impl to support this
- Update all examples to use updated API
- Tidied up ordering of graph_impl declarations a little

Author: Bensuo

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

@@ -53,7 +53,8 @@ class __SYCL_EXPORT node {
 template <graph_state State = graph_state::modifiable>
 class __SYCL_EXPORT command_graph {
 public:
-  command_graph(const property_list &propList = {});
+  command_graph(const context &syclContext, const device &syclDevice,
+                const property_list &propList = {});
 
   // Adding empty node with [0..n] predecessors:
   node add(const std::vector<node> &dep = {}) { return add_impl(dep); }
@@ -67,8 +68,7 @@ class __SYCL_EXPORT command_graph {
   void make_edge(node sender, node receiver);
 
   command_graph<graph_state::executable>
-  finalize(const sycl::context &syclContext,
-           const property_list &propList = {}) const;
+  finalize(const property_list &propList = {}) const;
 
   /// Change the state of a queue to be recording and associate this graph with
   /// it.
@@ -138,7 +138,6 @@ template <> class __SYCL_EXPORT command_graph<graph_state::executable> {
   void finalize_impl();
 
   int MTag;
-  const sycl::context &MCtx;
   std::shared_ptr<detail::exec_graph_impl> impl;
 };
 } // namespace experimental

sycl/source/detail/graph_impl.cpp

@@ -241,13 +241,12 @@ void exec_graph_impl::find_real_deps(std::vector<pi_ext_sync_point> &Deps,
   }
 }
 
-void exec_graph_impl::create_pi_command_buffers(sycl::device D,
-                                                const sycl::context &Ctx) {
+void exec_graph_impl::create_pi_command_buffers(sycl::device D) {
   // TODO we only have a single command-buffer per graph here, but
   // this will need to be multiple command-buffers for non-trivial graphs
   pi_ext_command_buffer OutCommandBuffer;
   pi_ext_command_buffer_desc Desc{};
-  auto ContextImpl = sycl::detail::getSyclObjImpl(Ctx);
+  auto ContextImpl = sycl::detail::getSyclObjImpl(MContext);
   const sycl::detail::plugin &Plugin = ContextImpl->getPlugin();
   auto DeviceImpl = sycl::detail::getSyclObjImpl(D);
   pi_result Res =
@@ -284,13 +283,13 @@ void exec_graph_impl::create_pi_command_buffers(sycl::device D,
                                           Node->MKernelName);
     }
 
-    auto SetFunc = [&Plugin, &PiKernel, &Ctx](sycl::detail::ArgDesc &Arg,
+    auto SetFunc = [&Plugin, &PiKernel, this](sycl::detail::ArgDesc &Arg,
                                               size_t NextTrueIndex) {
       sycl::detail::SetArgBasedOnType(
           Plugin, PiKernel,
           nullptr /* TODO: Handle spec constants and pass device image here */,
-          nullptr /* TODO: Pass getMemAllocation function for buffers */, Ctx,
-          false, Arg, NextTrueIndex);
+          nullptr /* TODO: Pass getMemAllocation function for buffers */,
+          this->MContext, false, Arg, NextTrueIndex);
     };
     std::vector<sycl::detail::ArgDesc> Args;
     sycl::detail::applyFuncOnFilteredArgs(EliminatedArgMask, Node->MArgs,
@@ -421,8 +420,9 @@ sycl::event exec_graph_impl::enqueue(
 
 template <>
 command_graph<graph_state::modifiable>::command_graph(
+    const sycl::context &syclContext, const sycl::device &syclDevice,
     const sycl::property_list &)
-    : impl(std::make_shared<detail::graph_impl>()) {}
+    : impl(std::make_shared<detail::graph_impl>(syclContext, syclDevice)) {}
 
 template <>
 node command_graph<graph_state::modifiable>::add_impl(
@@ -465,8 +465,9 @@ void command_graph<graph_state::modifiable>::make_edge(node Sender,
 template <>
 command_graph<graph_state::executable>
 command_graph<graph_state::modifiable>::finalize(
-    const sycl::context &CTX, const sycl::property_list &) const {
-  return command_graph<graph_state::executable>{this->impl, CTX};
+    const sycl::property_list &) const {
+  return command_graph<graph_state::executable>{this->impl,
+                                                this->impl->get_context()};
 }
 
 template <>
@@ -531,7 +532,7 @@ bool command_graph<graph_state::modifiable>::end_recording(
 
 command_graph<graph_state::executable>::command_graph(
     const std::shared_ptr<detail::graph_impl> &Graph, const sycl::context &Ctx)
-    : MTag(rand()), MCtx(Ctx),
+    : MTag(rand()),
       impl(std::make_shared<detail::exec_graph_impl>(Ctx, Graph)) {
   finalize_impl(); // Create backend representation for executable graph
 }
@@ -540,8 +541,8 @@ void command_graph<graph_state::executable>::finalize_impl() {
   // Create PI command-buffers for each device in the finalized context
   impl->schedule();
 #if SYCL_EXT_ONEAPI_GRAPH
-  for (auto device : MCtx.get_devices()) {
-    impl->create_pi_command_buffers(device, MCtx);
+  for (auto device : impl->get_context().get_devices()) {
+    impl->create_pi_command_buffers(device);
   }
 #endif
 }

sycl/source/detail/graph_impl.hpp

@@ -127,9 +127,10 @@ struct node_impl {
 };
 
 struct graph_impl {
-  std::set<std::shared_ptr<node_impl>> MRoots;
 
-  std::shared_ptr<graph_impl> MParent;
+  graph_impl(const sycl::context &syclContext, const sycl::device &syclDevice)
+      : MContext(syclContext), MDevice(syclDevice), MRecordingQueues(),
+        MEventsMap() {}
 
   void add_root(const std::shared_ptr<node_impl> &);
   void remove_root(const std::shared_ptr<node_impl> &);
@@ -155,8 +156,6 @@ struct graph_impl {
   std::shared_ptr<node_impl>
   add(const std::vector<std::shared_ptr<node_impl>> &Dep = {});
 
-  graph_impl() = default;
-
   /// Add a queue to the set of queues which are currently recording to this
   /// graph.
   void
@@ -199,8 +198,18 @@ struct graph_impl {
   /// an empty node is used to schedule dependencies on this sub graph.
   std::shared_ptr<node_impl>
   add_subgraph_nodes(const std::list<std::shared_ptr<node_impl>> &NodeList);
+  sycl::context get_context() const { return MContext; }
+
+  std::set<std::shared_ptr<node_impl>> MRoots;
+  std::shared_ptr<graph_impl> MParent;
 
 private:
+  // Context associated with this graph.
+  sycl::context MContext;
+  // Device associated with this graph. All graph nodes will execute on this
+  // device.
+  sycl::device MDevice;
+  // Unique set of queues which are currently recording to this graph.
   std::set<std::shared_ptr<sycl::detail::queue_impl>> MRecordingQueues;
   // Map of events to their associated recorded nodes.
   std::unordered_map<std::shared_ptr<sycl::detail::event_impl>,
@@ -224,7 +233,9 @@ class exec_graph_impl {
   sycl::event exec(const std::shared_ptr<sycl::detail::queue_impl> &);
   /// Turns our internal graph representation into PI command-buffers for a
   /// device
-  void create_pi_command_buffers(sycl::device D, const sycl::context &Ctx);
+  void create_pi_command_buffers(sycl::device D);
+
+  sycl::context get_context() const { return MContext; }
 
   const std::list<std::shared_ptr<node_impl>> &get_schedule() const {
     return MSchedule;

sycl/test/graph/graph-explicit-dotp-buffer.cpp

@@ -26,7 +26,8 @@ int main() {
 
   sycl::queue q{sycl::gpu_selector_v};
 
-  sycl::ext::oneapi::experimental::command_graph g;
+  sycl::ext::oneapi::experimental::command_graph g{q.get_context(),
+                                                   q.get_device()};
 
   float dotpData = 0.f;
   std::vector<float> xData(n);
@@ -93,7 +94,7 @@ int main() {
 #endif
     });
 
-    auto executable_graph = g.finalize(q.get_context());
+    auto executable_graph = g.finalize();
 
     // Using shortcut for executing a graph of commands
     q.ext_oneapi_graph(executable_graph).wait();

sycl/test/graph/graph-explicit-dotp-device-mem.cpp

@@ -26,7 +26,8 @@ int main() {
 
   sycl::queue q{sycl::gpu_selector_v};
 
-  sycl::ext::oneapi::experimental::command_graph g;
+  sycl::ext::oneapi::experimental::command_graph g{q.get_context(),
+                                                   q.get_device()};
 
   float *dotp = sycl::malloc_device<float>(1, q);
 
@@ -83,7 +84,7 @@ int main() {
       },
       {node_a, node_b});
 
-  auto executable_graph = g.finalize(q.get_context());
+  auto executable_graph = g.finalize();
 
   // Using shortcut for executing a graph of commands
   q.ext_oneapi_graph(executable_graph).wait();

sycl/test/graph/graph-explicit-dotp.cpp

@@ -26,7 +26,8 @@ int main() {
 
   sycl::queue q{sycl::gpu_selector_v};
 
-  sycl::ext::oneapi::experimental::command_graph g;
+  sycl::ext::oneapi::experimental::command_graph g{q.get_context(),
+                                                   q.get_device()};
 
   float *dotp = sycl::malloc_shared<float>(1, q);
 
@@ -83,7 +84,7 @@ int main() {
       },
       {node_a, node_b});
 
-  auto executable_graph = g.finalize(q.get_context());
+  auto executable_graph = g.finalize();
 
   // Using shortcut for executing a graph of commands
   q.ext_oneapi_graph(executable_graph).wait();

sycl/test/graph/graph-explicit-empty.cpp

@@ -7,7 +7,8 @@ int main() {
 
   sycl::queue q{sycl::gpu_selector_v};
 
-  sycl::ext::oneapi::experimental::command_graph g;
+  sycl::ext::oneapi::experimental::command_graph g{q.get_context(),
+                                                   q.get_device()};
 
   const size_t n = 10;
   float *arr = sycl::malloc_device<float>(n, q);
@@ -35,7 +36,7 @@ int main() {
       },
       {empty2});
 
-  auto executable_graph = g.finalize(q.get_context());
+  auto executable_graph = g.finalize();
 
   q.submit([&](sycl::handler &h) {
      h.ext_oneapi_graph(executable_graph);

sycl/test/graph/graph-explicit-multiple-exec-graphs.cpp

@@ -26,7 +26,8 @@ int main() {
 
   sycl::queue q{sycl::gpu_selector_v};
 
-  sycl::ext::oneapi::experimental::command_graph g;
+  sycl::ext::oneapi::experimental::command_graph g{q.get_context(),
+                                                   q.get_device()};
 
   float *dotp = sycl::malloc_shared<float>(1, q);
 
@@ -73,14 +74,14 @@ int main() {
       },
       {node_a, node_b});
 
-  auto executable_graph = g.finalize(q.get_context());
+  auto executable_graph = g.finalize();
 
   // Add an extra node for the second executable graph which modifies the output
   auto node_d =
       g.add([&](sycl::handler &h) { h.single_task([=]() { dotp[0] += 1; }); },
             {node_c});
 
-  auto executable_graph_2 = g.finalize(q.get_context());
+  auto executable_graph_2 = g.finalize();
 
   // Using shortcut for executing a graph of commands
   q.ext_oneapi_graph(executable_graph).wait();

sycl/test/graph/graph-explicit-node-ordering.cpp

@@ -7,7 +7,8 @@ int main() {
 
   sycl::queue q{sycl::gpu_selector_v};
 
-  sycl::ext::oneapi::experimental::command_graph g;
+  sycl::ext::oneapi::experimental::command_graph g{q.get_context(),
+                                                   q.get_device()};
 
   const size_t n = 10;
   float *x = sycl::malloc_shared<float>(n, q);
@@ -36,7 +37,7 @@ int main() {
   g.make_edge(init, mult);
   g.make_edge(mult, add);
 
-  auto executable_graph = g.finalize(q.get_context());
+  auto executable_graph = g.finalize();
 
   q.submit([&](sycl::handler &h) {
      h.ext_oneapi_graph(executable_graph);

sycl/test/graph/graph-explicit-queue-shortcuts.cpp

@@ -9,7 +9,8 @@ int main() {
 
   // Test passing empty property list, which is the default
   sycl::property_list empty_properties;
-  sycl::ext::oneapi::experimental::command_graph g(empty_properties);
+  sycl::ext::oneapi::experimental::command_graph g(
+      q.get_context(), q.get_device(), empty_properties);
 
   const size_t n = 10;
   float *arr = sycl::malloc_shared<float>(n, q);
@@ -21,7 +22,7 @@ int main() {
     });
   });
 
-  auto executable_graph = g.finalize(q.get_context(), empty_properties);
+  auto executable_graph = g.finalize(empty_properties);
 
   auto e1 = q.ext_oneapi_graph(executable_graph);
   auto e2 = q.ext_oneapi_graph(executable_graph, e1);