Incorrect error message when setting configure flag: --enable-nvtx-trace-per-dev

[ryanhankins]

When I run configure as follows:

./configure --prefix=/home/users/hankinsr/perf/aws-ofi-nccl/build --with-libfabric=/opt/cray/libfabric/1.15.2.0/ --with-hwloc=/home/users/hankinsr/perf/hwloc/build --with-cuda=/opt/nvidia/hpc_sdk/Linux_x86_64/23.11/cuda/12.3 --enable-nvtx-trace-per-dev --with-nvtx=/opt/nvidia/hpc_sdk/Linux_x86_64/23.3/cuda/12.0/targets/x86_64-linux

I get the following error message:

configure: error: Error: setting both nvtx_trace_per_comm and nvtx_trace_per_dev is currently not supported

But I'm only specifying -per-dev!

[rauteric]

Hi! Yes, we should improve this. nvtx_trace_per_comm currently always defaults to 1 (

aws-ofi-nccl/m4/check_pkg_nvtx.m4

Lines 41 to 42 in bacae0c

	[nvtx_trace_per_comm=1
	nvtx_trace_per_dev=0

), but we should make it default to 1 only when neither of the trace_per_comm nor trace_per_dev options is specified.

In the mean time, you can add --disable-nvtx-trace-per-comm.

[aws-nslick]

@rauteric is a build option the correct way to do this? nsight has flags which can filter events to specific domains/etc., it might be overall simpler to remove these build-time tunables and make it all-or-nothing on our end, but tag them on different domains and then provide instructions on how to only enable one or the other.