Nsight systems cuDNN support

Developer Tools Nsight Systems Profiling Linux Targets
1

Hi,
I am profiling a target x86 machine with CUDA 10.0 and cuDNN 7.6.1.34-1.

In profiles Timeline View I see no cuDNN events.
In Diagnostics Summary there is a message:
“Tracing libToolsInjectionCuDNN64_7_6.so library version 7.6 is currently not supported, library name: … .local/share/nsight_systems/libToolsInjectionCuDNN64_7_6.so, error: … .local/share/nsight_systems/libToolsInjectionCuDNN64_7_6.so: cannot open shared object file: No such file or directory”

And indeed, I have only .so files for the latest cudnn version being libToolsInjectionCuDNN64_7_4.so.

Does Nsight Systems support cuDNN 7.5 and 7.6?

Regards,
Peter

2

What version of Nsight Systems are you using?

3

I was using 2019.2.1 GUI on Mac OS to remotely profile applications on a Linux host.

Now I am trying to use the latest version - 2019.4.

Installing on a Linux host with NVDRV:410.78, CUDA:10.0, cuDNN:7.6.1.34-1.

Installed CLI-only package with:
$ sudo dpkg -i NVIDIA_Nsight_Systems_Linux_CLI_Only_2019.4.2.135.deb

Now I always get segmentation faults.

$ nsys profile python cifar100_VGG.py -b 300 -e 1 --iterations 1

**** collection configuration ****
	force-overwrite = false
	stop-on-exit = true
	export_sqlite = false
	stats = false
	capture-range = none
	stop-on-range-end = false
	Beta: ftrace events:
	ftrace-keep-user-config = false
	delay = 0 seconds
	duration = 0 seconds
	kill = none
	inherit-environment = true
	show-output = true
	trace-fork-before-exec = false
	sample_cpu = true
	backtrace_method = LBR
	wait = all
	trace_cublas = false
	trace_cuda = true
	trace_cudnn = false
	trace_nvtx = true
	trace_mpi = false
	trace_openacc = false
	trace_vulkan = false
	trace_opengl = true
	trace_osrt = true
	osrt-threshold = 0 nanoseconds
	cudabacktrace = false
	cudabacktrace-threshold = 0 nanoseconds
	profile_processes = tree
	application command = python
	application arguments = cifar100_VGG.py -b 300 -e 1 --iterations 1
	application working directory = /HDD2/ML/mlbenchmarks/Keras/keras
	NVTX profiler range trigger =
	NVTX profiler domain trigger =
	environment variables:
Segmentation fault (core dumped)
4

So, there is a bug in 2019.4.2.135, and I sent off a request for an update of the downloadable file to 2019.4.2.139 last week. I’m glad you reported this, because now I see that that update did not happen. I am working on fixing that. Meanwhile, if you email me directly at hwilper@nvidia.com, I will get you the correct bits.