Is my understanding correct that CLOffset represents the word offset? @rahulbera

[szuhzw]

[rahulbera]

HI @szuhzw. You are correct. This is a bug. Technically, generate_index_from_feature function should call process_CLOffset (or process_PC_CLOffset) for CLOffset (or PC_CLOffset) feature type. But it was mistakenly calling the process_CLWordOffset. The same bug is also present for CLDWordOffset feature too. This bug likely got in during refactoing the code for GitHub release.

I am creating a patch to fix this issue. Thanks for pointing it out.

[rahulbera]

Hi @szuhzw, I merged this pull request, which should fix this issue.

Now that this bug has been caught and fixed, can you please check and report what's the performance difference (in terms of accuracy, coverage, and overall IPC) of Hermes across PC_CLOffset, PC_CLWordOffset, PC_CLDWordOffset features? Thanks.

[szuhzw]

Hi @rahulbera , The latest perc_pred_helper.cc doesn't define the process_PC_CLDWordOffset function.

[szuhzw]

Hi @rahulbera ,
uint32_t process_PC_CLDwordOffset(state_info_t *state, uint64_t metadata, int32_t hash_type, uint32_t weight_array_size).
Is it right that the process_PC_CLDwordOffset should be process_PC_CLDWordOffset?

[szuhzw]

Hi. @rahulbera , Do you mean you want me to test the performance of the following three sets of features?
5,8,9,11,16
5,8,9,13,16
5,8,9,15,16

[rahulbera]

Hi @szuhzw, yes. There was a typo in the function name process_PC_CLDwordOffset. It's fixed now.

[rahulbera]

Regarding your second question: yes. Since the bug is fixed now, we should do a regression test with the three feature sets that you mentioned.

[szuhzw]

Hi. @rahulbera
I ran over 200 traces. In terms of accuracy, the difference between “5,8,9,11,16” and "5,8,9,13,16" is extremely small, while "5,8,9,15,16"'s accuracy is about 2% lower than “5,8,9,11,16” and "5,8,9,13,16". In terms of recall, the difference in coverage among these three sets of features is extremely small. In terms of single-core performance, the geometric mean IPC difference between these three feature sets is extremely small.

[rahulbera]

Thanks @szuhzw for running the regression. Good to know that the results are not differing much.

[szuhzw]

Hi. @rahulbera
Has the research group you are in ever reviewed an paper titled "*****: An Ultra-Lightweight Level High-Performance Off-Chip Load Predictor for Long-Latency Load Requests"?