I have frequent blanck screens happening when I reach specific web pages (such as amazon prime), on multiple browsers.
I wrote to the reseller (LDLC) to get support, but until I get support, I thought about profiling what is happening with nsight systems.
It appears that the card is becoming idle, following a thread alert.
The idle state is manifested by a blanck screen but it seems that, because of latencies to go off and on, there are more frequent thread alerts. In other words, a black screen can actually be covered by multiple thread alerts.
Do you have pieces of advice to investigate more ? like getting a call stack and finding the culprit.
And how to get in contact with nvidia to share the report so a correction of the drivers can be implemented ?
I use nv driver 572.47 and windows 10 enterprise 19045.5487
@hwilper are those nvidia forums ? do nvidia people from driver teams interact in those ? About the calls stacks, with intel adviser it is possible to get some, but they are not related to a moment (specific or selected range), only from the whole sampling, so I won’t be able to know what is happening during those idle moment; and to my knowledge, I cannot connect the nsight systems to adviser. Could procdump64 be helpful ?
@Gatchan those are NVIDIA forums and they should have the driver team interacting there.
With Nsight Systems you should be able to set a range with NVTX annotations and then you will be able to know what backtraces are inside that range. We do not currently have any direct data path from Nsight Systems to Intel Advisor.
In the report, I asked NVTX, but “stats system view” tells me the “report (report1.sqlite) does not contain NV Tools Extension (NVTX) data”. I did not put any filter, I did not ask to insert any marker with a hotkey, should I, to get something ?
I tried to add NVTX markers with the hotkey, it aborts the blocking thread and the idel mode of the gpu but I could not see the markers in the profile result.
I am not sure if the annotation is the same as NVTX markers.
Hi,
The wait reason that appears on the thread state row “Alert by Thread Id” is just the Windows kernel way of saying that the thread is blocked on a synchronization event, such as waiting to take ownership of a mutex.
If you enabled CPU Context Switch backtrace collection in the project settings (from the screenshot it looks like you did) then you can hover the mouse over the blank range in the thread execution row and see a backtrace of the callstack that the thread has executed when the blocking wait wait call was made. Use the Resolve Symbols option and provide the Microsoft Public Symbol Server URL and the location of your debug info (*.pdb) files. Nsight Systems will replace the callstack IP addresses with symbol names (i.e. function names). This will provide more context about the operation the thread was trying to perform when it blocked on the sync call.
I suggest to expand the GPU row at the top of the report timeline to expose the VSync row. Is this renderer waiting for a VSync event?
