igt/gem_exec_nop: Relax assertion for parallel execution
In an ideal world, we should be able to execute on every engine in
parallel and the single limiting factor would be how fast the GPU can
execute. Due to the serialisation in execbuf, we would lockstep with
execution to the slowest engine and so would execute the same number of
cycles on each. However in CI, we are limited by how fast the driver is,
particularly under invasive debugging. This makes asserting that the
average time == max/nengine impossible, and reveals that the assertion is
impossible to meet under general condition. It's an impractical
regression test. Therefore we relax the assertion to only detect should
something critically fail. Worst case behaviour is presumed that each
ring runs sequentially, and so running N rings in parallel should take no
longer than running N rings serially. (Pathologically it can be even
slower if no batching on the rings occur).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
diff --git a/tests/gem_exec_nop.c b/tests/gem_exec_nop.c
index 9b89260..480eb8f 100644
--- a/tests/gem_exec_nop.c
+++ b/tests/gem_exec_nop.c
@@ -184,17 +184,25 @@
igt_assert_eq(intel_detect_and_clear_missed_interrupts(fd), 0);
time = elapsed(&start, &now) / count;
- igt_info("All (%d engines): %'lu cycles, average %.3fus per cycle\n",
- nengine, count, 1e6*time);
+ igt_info("All (%d engines): %'lu cycles, average %.3fus per cycle [expected ideal %.3fus]\n",
+ nengine, count, 1e6*time, 1e6*max/nengine);
- /* The rate limiting step is how fast the slowest engine can
- * its queue of requests, if we wait upon a full ring all dispatch
- * is frozen. So in general we cannot go faster than the slowest
- * engine, but we should equally not go any slower.
+ /* The rate limiting step should be how fast the slowest engine can
+ * execute its queue of requests, as when we wait upon a full ring all
+ * dispatch is frozen. So in general we cannot go faster than the
+ * slowest engine (but as all engines are in lockstep, they should all
+ * be executing in parallel and so the average should be max/nengines),
+ * but we should equally not go any slower.
+ *
+ * However, that depends upon being able to submit fast enough, and
+ * that in turns depends upon debugging turned off and no bottlenecks
+ * within the driver. We cannot assert that we hit ideal conditions
+ * across all engines, so we only look for an outrageous error
+ * condition.
*/
- igt_assert_f(time < max + 10*min/9, /* ensure parallel execution */
- "Average time (%.3fus) exceeds expecation for parallel execution (min %.3fus, max %.3fus; limit set at %.3fus)\n",
- 1e6*time, 1e6*min, 1e6*max, 1e6*(max + 10*min/9));
+ igt_assert_f(time < 2*sum,
+ "Average time (%.3fus) exceeds expectation for parallel execution (min %.3fus, max %.3fus; limit set at %.3fus)\n",
+ 1e6*time, 1e6*min, 1e6*max, 1e6*sum*2);
}
static void print_welcome(int fd)