Icemist commented on code in PR #13726:
URL: https://github.com/apache/tvm/pull/13726#discussion_r1064139675
##########
src/runtime/profiling.cc:
##########
@@ -894,15 +891,21 @@ PackedFunc WrapTimeEvaluator(PackedFunc pf, Device dev,
int number, int repeat,
std::max((min_repeat_ms / (duration_ms / number) + 1), number *
golden_ratio));
}
+ int64_t accum_t_nanos = 0;
// start timing
- Timer t = Timer::Start(dev);
for (int j = 0; j < number; ++j) {
+ // call preprocessing function
+ if (f_preproc != nullptr) {
+ f_preproc.CallPacked(args, &temp);
Review Comment:
From experience that this change needs to be reversed.
The **repeat** and **number** parameters are designed for different
scenarios. For your scenario, you need to increase the **repeat** and use
**number=1**.
Scenario with increasing **number** (or using **min_repeat_ms**) is used
when it is problematic to measure single runs - "_The timer is specifically in
the outer loop so that we can time functions with a very small runtime or use
less precise timers._"
--
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]
