benchdnn: cold cache improvement #4529
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
abort() prevents singleton destructions which print misleading and anooying messages in case of incorrectly finished application due to their nature to catch errors for correctly finished one. It also allows to catch errors on the spot with gdb instead of breaking on a specific line from the message.
Timer can now identify spikes in collection results and discard them when reporting final results. This is done to stabilize best statistics for GPU perf validation.
github-actions
bot
added
platform:gpu-intel
Contributor
Author
|
make test perf-gpu |
echeresh
reviewed
Jan 13, 2026
| constexpr double magnitude_threshold = 1.1; | ||
| size_t major_magnitude = SIZE_MAX; | ||
| for (size_t i = 0; i < deltas.size(); i++) { | ||
| deltas[i] = ms_vec_[i + 1] / ms_vec_[i]; |
Contributor
There was a problem hiding this comment.
Do we need to check bounds? Say, if ms_vec_.size() == 1 and deltas_size == 1, this goes out of bounds.
Contributor
There was a problem hiding this comment.
Another comment: do we really need deltas vector? Looks like its values are not used outside of the loop.
| } | ||
| } | ||
|
|
||
| if (major_magnitude < deltas.size()) { |
Contributor
There was a problem hiding this comment.
Suggested change
| if (major_magnitude < deltas.size()) { | |
| if (major_magnitude < deltas_size) { |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
LNL may report higher BW than HW provides for certain shapes (the last number is expected to be <= 108):
The issue lies in the hardware behavior when in certain situations L3 cache won't be flushed and input memories will reside in there even for a cold-cache mode.
The resolution is to introduce a flushing kernel that is submitted after each execution for trigger that L3 cache flushing. The credit for idea. library implementation and pivoting of PoC goes to @echeresh.
Note: even though the flushing kernel does stabilize result significantly, it may happen there's a one or two hardware "shots" where the number would still be higher as reported above which is caused by unknown effect at this moment.
The work around those "shots" is to drop several fastest results from the timer collection
Finally, flushing kernel costs time but with better stability time which allows to decrease the sample size.
Three changes combined should address the "not-cold-enough" cold-cache issue.