Hello everybody,
is it possible to somehow access the hardware performance counters of the GPU on the Jetson TX2 platform? I would like to analyze memory accesses on this low level.
Best regards and thank you in advance!
Hello everybody,
is it possible to somehow access the hardware performance counters of the GPU on the Jetson TX2 platform? I would like to analyze memory accesses on this low level.
Best regards and thank you in advance!
Hi,
You can get some hardware execution detail with nvprof.
Would you mind to check if the profiling data is sufficient for you first?
[url]https://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvprof-overview[/url]
Thanks.
Hello,
thanks for your answer. Unfortunately the information given by nvprof is not enough. This post states that somehow cache hit analysis is not available on Jetson TX2:
https://devtalk.nvidia.com/default/topic/1047884/profile-tx2-cache-misses/
I saw this list of performance counters accessible with Tegra Graphics Debugger:
https://docs.nvidia.com/gameworks/index.html#developertools/mobile/tegra_graphics_debugger/tgd_perf_counters.htm%3FTocPath%3DDeveloper%2520Tools|Mobile%2520Developer%2520Tools|Tegra%2520Graphics%2520Debugger|_____8
Are those also valid for Jetson TX2 and can I access them to profile CUDA applications?
Hi,
Suppose not.
Since the way to profile cache miss for ARM is quite different.
But you still can give it a try to check if any useful information.
Thanks.