Error in read_reports

[Diksha-Moolchandani]

Hi,

I get the following error in the last step of the README:

drgpu/read_reports.py", in fill_source_report
new_dict = {line_id: source_df.at[line_id, stall_reason]}
pandas/core/indexing.py", line 2488, in getitem
return super().getitem(key)
pandas/core/indexing.py", line 2440, in getitem
return self.obj._get_value(*key, takeable=self._takeable)
pandas/core/frame.py", line 4008, in _get_value
series = self._get_item_cache(col)
pandas/core/frame.py", line 4417, in _get_item_cache
loc = self.columns.get_loc(item)
pandas/core/indexes/base.py", line 3797, in get_loc
raise KeyError(key) from err
KeyError: 'stall_barrier'

Have you seen this error? What might be the reason?

[FindHao]

It looks like the source.csv you exported doesn't have the stall_barrier column. can you double check it?

[Diksha-Moolchandani]

Which file should contain stall_barrier? btree.csv or btree_s.csv?

btree.csv has some counters related to stalls but btree_s.csv has the source code with some annotations.

Can you share a sample of your both csv files?

[FindHao]

I don't have proper servers with V100 and GTX1650. But I'll find some time to get some reports from other machines to reproduce your errors. If you want, you can also share your csv files, and I'll take a look at the bugs.

[Diksha-Moolchandani]

btree_s.csv
btree.csv

These are the 2 files. Please have a look and let me know

[FindHao]

Hi,

I just found some previous log files. Here is the btree_s.csv I collected a long time ago. It looks like ncu doesn't add stall-related metrics in exported source code anymore. I'll check if there are configs to bring it back and how the stall-related metrics are handled in source code mode.
btree_source.csv

[Diksha-Moolchandani]

Any update on the issue?

[FindHao]

Sorry for the late reply. This view is hided in ncu-ui by default. You can open report file in ncu-ui. Go to the source page --> source view --> navigation --> more.

check the Warp Stall group.

Then export the current page to csv file. I didn't find a command line solution, so you have to do it manually for now.

[Diksha-Moolchandani]

This does not appear for me. There is no such option as Warp stall. Which version of nsight do you have?

[FindHao]

I used the latest NCU.