[SYCL] Update to execute graph through the handler

Will require reble/llvm#42 to be encorporated
into the `sycl-graphs-record-and-replay` branch to work

Also added a new test `graph_record_queue_shortcuts.cpp` to test the
various graph execution queue shortcuts.

Author: EwanC

SYCL/CommandGraph/graph_finalize_empty.cpp

@@ -18,8 +18,8 @@ int main() {
       graph;
   auto graphExec = graph.finalize(testQueue.get_context());
 
-  testQueue.submit(graphExec);
+  testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
   testQueue.wait();
 
   return 0;
-}
+}

SYCL/CommandGraph/graph_record_after_finalize.cpp

@@ -73,11 +73,11 @@ int main() {
 
     // Execute several iterations of the graph
     for (unsigned n = 0; n < iterations; n++) {
-      testQueue.submit(graphExec);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
     }
     // Execute the extended graph.
     for (unsigned n = 0; n < iterations; n++) {
-      testQueue.submit(graphExecAdditional);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExecAdditional); });
     }
     // Perform a wait on all graph submissions.
     testQueue.wait_and_throw();
@@ -88,4 +88,4 @@ int main() {
   failed |= referenceOut != dataOut;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_after_use.cpp

@@ -50,7 +50,7 @@ int main() {
     // Execute several iterations of the graph (first iteration has already run
     // before graph recording)
     for (unsigned n = 1; n < iterations; n++) {
-      testQueue.submit(graphExec);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
     }
     // Perform a wait on all graph submissions.
     testQueue.wait();
@@ -62,4 +62,4 @@ int main() {
   failed |= referenceC != dataC;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_basic.cpp

@@ -46,7 +46,7 @@ int main() {
 
     // Execute several iterations of the graph
     for (unsigned n = 0; n < iterations; n++) {
-      testQueue.submit(graphExec);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
     }
     // Perform a wait on all graph submissions.
     testQueue.wait();
@@ -58,4 +58,4 @@ int main() {
   failed |= referenceC != dataC;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_basic_usm.cpp

@@ -49,7 +49,7 @@ int main() {
 
     // Execute several iterations of the graph
     for (unsigned n = 0; n < iterations; n++) {
-      testQueue.submit(graphExec);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
     }
     // Perform a wait on all graph submissions.
     testQueue.wait();
@@ -69,4 +69,4 @@ int main() {
   failed |= referenceC != dataC;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_double_buffer.cpp

@@ -60,10 +60,10 @@ int main() {
     testQueue.end_recording();
 
     for (size_t i = 0; i < iterations; i++) {
-      testQueue.submit(execGraph);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
       // Update to second set of buffers
       execGraph.update(graphUpdate);
-      testQueue.submit(execGraph);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
       // Reset back to original buffers
       execGraph.update(graph);
     }
@@ -81,4 +81,4 @@ int main() {
   failed |= referenceC2 != dataC2;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_host_task.cpp

@@ -78,7 +78,7 @@ int main() {
 
     // Execute several iterations of the graph
     for (unsigned n = 0; n < iterations; n++) {
-      testQueue.submit(graphExec);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
     }
     // Perform a wait on all graph submissions.
     testQueue.wait();
@@ -88,4 +88,4 @@ int main() {
   failed |= referenceC != dataC;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_multiple_exec_graphs.cpp

@@ -50,7 +50,7 @@ int main() {
     // before graph recording)
     for (unsigned n = 1; n < iterations; n++) {
       auto graphExec = graph.finalize(testQueue.get_context());
-      testQueue.submit(graphExec);
+      testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
     }
     // Perform a wait on all graph submissions.
     testQueue.wait();
@@ -62,4 +62,4 @@ int main() {
   failed |= referenceC != dataC;
 
   return failed;
-}
+}

SYCL/CommandGraph/graph_record_queue_shortcuts.cpp

@@ -0,0 +1,64 @@
+// REQUIRES: level_zero, gpu
+// RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
+// RUN: %GPU_RUN_PLACEHOLDER %t.out
+
+/** Tests queue shortcuts for executing a graph */
+
+#include "graph_common.hpp"
+
+using namespace sycl;
+
+int main() {
+  queue testQueue;
+
+  using T = int;
+
+  std::vector<T> dataA(size), dataB(size), dataC(size);
+
+  // Initialize the data
+  std::iota(dataA.begin(), dataA.end(), 1);
+  std::iota(dataB.begin(), dataB.end(), 10);
+  std::iota(dataC.begin(), dataC.end(), 1000);
+
+  // Create reference data for output
+  std::vector<T> referenceA(dataA), referenceB(dataB), referenceC(dataC);
+  calculate_reference_data(iterations, size, referenceA, referenceB,
+                           referenceC);
+
+  {
+    ext::oneapi::experimental::command_graph<
+        ext::oneapi::experimental::graph_state::modifiable>
+        graph;
+    buffer<T> bufferA{dataA.data(), range<1>{dataA.size()}};
+    buffer<T> bufferB{dataB.data(), range<1>{dataB.size()}};
+    buffer<T> bufferC{dataC.data(), range<1>{dataC.size()}};
+
+    testQueue.begin_recording(graph);
+
+    // Record commands to graph
+
+    run_kernels(testQueue, size, bufferA, bufferB, bufferC);
+
+    testQueue.end_recording();
+    auto graphExec = graph.finalize(testQueue.get_context());
+
+    // Execute several iterations of the graph using the different shortcuts
+    event e = testQueue.exec_graph(graphExec);
+
+    assert(iterations > 2);
+    const unsigned loop_iterations = iterations - 2;
+    std::vector<event> events(loop_iterations);
+    for (unsigned n = 0; n < loop_iterations; n++) {
+      events[n] = testQueue.exec_graph(graphExec, e);
+    }
+
+    testQueue.exec_graph(graphExec, events).wait();
+  }
+
+  bool failed = false;
+  failed |= referenceA != dataA;
+  failed |= referenceB != dataB;
+  failed |= referenceC != dataC;
+
+  return failed;
+}

SYCL/CommandGraph/graph_record_stream.cpp

@@ -44,11 +44,11 @@ int main() {
 
     auto graphExec = graph.finalize(testQueue.get_context());
 
-    testQueue.submit(graphExec);
+    testQueue.submit([&](handler &cgh) { cgh.exec_graph(graphExec); });
 
     // Perform a wait on all graph submissions.
     testQueue.wait();
   }
 
   return 0;
-}
+}