Measuring double-precision operations with DP_OPS event
219 views
Skip to first unread message
Fotis Drakopoulos
unread,
Apr 23, 2021, 10:07:32 AM4/23/21
Reply to author
Sign in to reply to author
Forward
Sign in to forward
Delete
You do not have permission to delete messages in this group
Copy link
Report message
Show original message
Either email addresses are anonymous for this group or you need the view member email addresses permission to view the original message
to ptools-...@icl.utk.edu
Dear,
I'm trying to use the PAPI software to measure the operations of a neural-network model in python using a python wrapper (pypapi). I'm using the event PAPI_DP_OPS (since the PAPI_FP_OPS event is not available on my CPU) and, when I'm counting an iteration of the neural-network model using the
PAPI_DP_OPS
event, I get around 390 FLOPS as output. This number seems rather low for the operations of a neural-network model, so I was wondering whether the
PAPI_DP_OPS event returns the total count of DP_OPS or maybe the MFLOPS/s (or something like that) that justifies this low number?
Thank you in advance!
Best regards,
Fotis
Anthony Danalis
unread,
Apr 23, 2021, 10:23:29 AM4/23/21
Reply to author
Sign in to reply to author
Forward
Sign in to forward
Delete
You do not have permission to delete messages in this group
Copy link
Report message
Show original message
Either email addresses are anonymous for this group or you need the view member email addresses permission to view the original message
to Fotis Drakopoulos, ptools-perfapi
Foti, PAPI_DP_OPS should return a count, not a rate. A way to verify
this on your machine would be to run:
papi_avail -e PAPI_DP_OPS
which will show you information about the event, and specifically the
native event(s) used to derive it.
Why do you expect the model to have multiple double precision floating
point operations? Typically, NN codes do not use many DP float
operations.
You do not have permission to delete messages in this group
Copy link
Report message
Show original message
Either email addresses are anonymous for this group or you need the view member email addresses permission to view the original message
to Anthony Danalis, ptools-perfapi
Hi Anthony,
Thanks for the answer, that's what I expected as well. My assumption was
that the NN model that I evaluated required a large amount of FLOPS, but
when I run the PAPI_SP_OPS I get 0 operations and ~390 for the
PAPI_DP_OPS. The reported values that I normally see for FLOPS for such
NN models are in the order of GFLOPS, are these reported operations
different than the DP_OPS that I get from papi?
Regards,
Fotis
Muhammet Balcilar
unread,
Jun 27, 2022, 8:57:14 PM6/27/22
Reply to author
Sign in to reply to author
Forward
Sign in to forward
Delete
You do not have permission to delete messages in this group
Copy link
Report message
Show original message
Either email addresses are anonymous for this group or you need the view member email addresses permission to view the original message
to ptools-perfapi, Fotis Drakopoulos, ptools-perfapi, adanalis
Probably you use GPU. PAPI daes not count GPU operations. Use just cpu and count
PAPI_SP_OPS
Anthony Danalis
unread,
Jun 28, 2022, 9:55:25 AM6/28/22
Reply to author
Sign in to reply to author
Forward
Sign in to forward
Delete
You do not have permission to delete messages in this group
Copy link
Report message
Show original message
Either email addresses are anonymous for this group or you need the view member email addresses permission to view the original message
to Muhammet Balcilar, ptools-perfapi, Fotis Drakopoulos
This statement is false. PAPI can count GPU operations on GPUs from
all three vendors (NVIDIA, AMD, and Intel). However, not by using the
event PAPI_SP_OPS. For GPU events to be counted the PAPI user has to
configure in the corresponding component and use the corresponding
native events, which can be found using the utility papi_native_avail.