Changes to test_async_submit

Few changes to reduce sporadic test failures.

  1. Choose initial global work-size to be a multiple of
     max_work_group_size to ensure that enough work is
     being generated in the test

  2. Retry submission if asyncronicity is not yet detected
     and double the work if that does not help

Author: oleksandr-pavlyk

dpctl/tests/test_sycl_kernel_submit.py

@@ -167,18 +167,21 @@ def test_async_submit():
     assert isinstance(kern2Kernel, dpctl_prog.SyclKernel)
 
     status_complete = dpctl.event_status_type.complete
+
     # choose input size based on capability of the device
     f = q.sycl_device.max_work_group_size
     n = f * 1024
-    X = dpt.empty((3, n), dtype="u4", usm_type="device", sycl_queue=q)
+    n_alloc = 4 * n
+
+    X = dpt.empty((3, n_alloc), dtype="u4", usm_type="device", sycl_queue=q)
     first_row = dpctl_mem.as_usm_memory(X[0])
     second_row = dpctl_mem.as_usm_memory(X[1])
     third_row = dpctl_mem.as_usm_memory(X[2])
 
     p1, p2 = 17, 27
 
     async_detected = False
-    for _ in range(5):
+    for attemp in range(5):
         e1 = q.submit(
             kern1Kernel,
             [
@@ -224,6 +227,10 @@ def test_async_submit():
             async_detected = True
             e3.wait()
             break
+        else
+            n = n * (1 if attempt % 2 == 0 else 2)
+            if n > n_alloc:
+                break
 
     assert async_detected, "No evidence of async submission detected, unlucky?"
     Xnp = dpt.asnumpy(X)