[SYCL] test for barrier batching (intel#750)

Author: smaslov-intel

SYCL/Plugin/level_zero_batch_barrier.cpp

@@ -0,0 +1,50 @@
+// REQUIRES: gpu, level_zero
+
+// RUN: %clangxx -fsycl -fsycl-unnamed-lambda -fsycl-targets=%sycl_triple %level_zero_options %s -o %t.out
+// RUN: env SYCL_PI_TRACE=2 ZE_DEBUG=1 %GPU_RUN_PLACEHOLDER %t.out 2>&1 %GPU_CHECK_PLACEHOLDER
+
+// Test that the wait with a barrier is fully batched, i.e. it doesn't cause
+// extra submissions.
+
+#include <CL/sycl.hpp>
+#include <level_zero/ze_api.h>
+#include <vector>
+
+using namespace std;
+using namespace sycl;
+
+void submit_kernel(queue &q) {
+  q.submit([&](auto &h) { h.parallel_for(1, [=](size_t id) {}); });
+}
+
+int main(int argc, char *argv[]) {
+  queue q;
+
+  submit_kernel(q); // this one will immediatelly execute since q is empty
+                    // CHECK: ---> piEnqueueKernelLaunch
+                    // CHECK: ZE ---> zeCommandQueueExecuteCommandLists
+
+  submit_kernel(q); // starts a batch
+                    // CHECK: ---> piEnqueueKernelLaunch
+                    // CHECK-NOT: ZE ---> zeCommandQueueExecuteCommandLists
+
+  // continue the batch
+  event barrier = q.ext_oneapi_submit_barrier();
+  // CHECK: ---> piEnqueueEventsWaitWithBarrier
+  // CHECK-NOT: ZE ---> zeCommandQueueExecuteCommandLists
+
+  submit_kernel(q);
+  // CHECK: ---> piEnqueueKernelLaunch
+  // CHECK-NOT: ZE ---> zeCommandQueueExecuteCommandLists
+
+  // interop should close the batch
+  ze_event_handle_t ze_event =
+      get_native<backend::ext_oneapi_level_zero>(barrier);
+  // CHECK: ---> piextEventGetNativeHandle
+  // CHECK: ZE ---> zeCommandQueueExecuteCommandLists
+  zeEventHostSynchronize(ze_event, UINT64_MAX);
+
+  // CHECK: ---> piQueueFinish
+  q.wait_and_throw();
+  return 0;
+}