performance profiling #675

hw-ju · 2022-04-27T20:15:34Z

hw-ju
Apr 27, 2022

Hi! Are with Range("XXX"): and with nvtx.annotate("XXX"): essentially the same thing?
Thanks!

Thanks for your question. They are not exactly the same but very similar. They both define the span of an event but there are two differences:

with Range uses torch.cuda.nvtx and it only depends on the cuda version of PyTorch. However, with nvtx.annotate uses nvtx library directly, so it depends on nvtx being installed.
with Range uses range_push and range_pop on the backend to define the context, so it inherit their limitations like they currently does not support to change the range color, while in with nvtx.annotate you can set the color in the input argument.

Nic-Ma · 2022-04-28T01:46:49Z

Could you please help share some comments about this question?

Thanks in advance.

0 replies

drbeh · 2022-04-28T12:42:17Z

Thanks for your question. They are not exactly the same but very similar. They both define the span of an event but there are two differences:

with Range uses torch.cuda.nvtx and it only depends on the cuda version of PyTorch. However, with nvtx.annotate uses nvtx library directly, so it depends on nvtx being installed.
with Range uses range_push and range_pop on the backend to define the context, so it inherit their limitations like they currently does not support to change the range color, while in with nvtx.annotate you can set the color in the input argument.

1 reply

Very helpful! Thanks @drbeh! :D