masahi commented on code in PR #10937:
URL: https://github.com/apache/tvm/pull/10937#discussion_r846944453
##########
src/meta_schedule/task_scheduler/task_scheduler.cc:
##########
@@ -127,7 +128,20 @@ void TaskSchedulerNode::Tune() {
ICHECK(!task->is_terminated);
ICHECK(!task->runner_futures.defined());
SearchStrategy strategy = task->search_strategy.value();
- if ((task->measure_candidates =
strategy->GenerateMeasureCandidates()).defined()) {
+ Optional<Array<MeasureCandidate>> candidates =
strategy->GenerateMeasureCandidates();
+ if (candidates.defined()) {
+ // Filter out invalid candidates.
+ // An invalid candidate can arise due to a schedule application failure
(e.g. tensorize).
+ Array<MeasureCandidate> valid_candidates;
+ for (MeasureCandidate candidate : candidates.value()) {
+ if (candidate.defined()) {
+ valid_candidates.push_back(candidate);
+ }
+ }
+
+ if (valid_candidates.size() == 0) continue;
Review Comment:
Yeah, I realized that after I put try/catch in `ThreadedApply`, all
candidates are valid if ` support::parallel_for_dynamic` ever returns.
The issue I have now is that, that infinite loop can now really repeat
forever, if some postprocs never succeed on all design spaces. I'm now trying
conv2d VNNI auto-tensorize, and due to an implementation issue in
auto-tensorize, `sch->Tensorize` in postproc always fail. But since the
tensorize error is now caught by `ThreadedApply` and we repeat, it leads to an
infinite loop.
So I'm now wondering if doing try/catch in `ThreadedApply` is the right
solution here. Rather, we might want to wrap try/catch around
`sch->Tensorize(...)` in a postproc rewrite, meaning we return non-tensorized
schedule in case of auto-tensorization failure. I think this is what @vinx13
suggested last week.
So, if we agree that we should wrap try/catch around `sch->Tensorize(...)`,
I don't need the change in this PR anymore. Do we have other good reason to put
try/catch around `ThreadedApply`? Otherwise I'll close this.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
