Misaligned double atomic, but only when NSight is debugging remotely
I'm doing a simple atomicAdd using doubles (win7/vs2015sp3/nsight 5.3)), with NSight debugging onto a remote machine. The remote card (1060) has 6.1 capability, which should be fine with a double atomicAdd. However NSight complains with a misaligned atomic. BTW, during the run, I requested the card's name, and its the right one. But if I *locally* run the exact same code on the (previously) remote machine, all's fine. Whats happening here?
I'm doing a simple atomicAdd using doubles (win7/vs2015sp3/nsight 5.3)), with NSight debugging onto a remote machine. The remote card (1060) has 6.1 capability, which should be fine with a double atomicAdd. However NSight complains with a misaligned atomic. BTW, during the run, I requested the card's name, and its the right one. But if I *locally* run the exact same code on the (previously) remote machine, all's fine. Whats happening here?

#1
Posted 07/21/2017 04:00 AM   
misaligned atomic, this error comes from nsight memcheck, could you have a check when the nsight memcheck is off, also could you tell me which driver you use.
misaligned atomic, this error comes from nsight memcheck, could you have a check when the nsight memcheck is off, also could you tell me which driver you use.

#2
Posted 07/24/2017 02:47 AM   
Thanks for replying.... If I uncheck the memory checker option under NSight, then there's no error notifications. Video driver is 384.76.
Thanks for replying....
If I uncheck the memory checker option under NSight, then there's no error notifications.
Video driver is 384.76.

#3
Posted 07/24/2017 03:31 AM   
Em, please check if your code has memory leak issue, I will try to repo it locally to identify if it's a bug. Also could you please use cuda-memcheck not the memcheck in nsight to check your app and find out the difference?
Em, please check if your code has memory leak issue, I will try to repo it locally to identify if it's a bug.

Also could you please use cuda-memcheck not the memcheck in nsight to check your app and find out the difference?

#4
Posted 07/24/2017 03:34 AM   
So I've simplified everything down to just NSight running locally on one machine, with one 1060 running the minimal code below. With command line cuda-memcheck, all's fine. The only issues I have is when the VS inbuilt NSight (running locally) has memory checker enabled. Also if I NSight to a remote machine, the behaviour remains. No memory fails too, as you can see. __global__ void why( double *dd, float *ff ) { float thingFloat = 3.0f; float qf = atomicAdd(ff, thingFloat); double thingDouble = 3.0f; atomicAdd(dd, thingDouble); } int main() { double *dd; cudaMalloc((void **)&dd, 100 * sizeof(double)); float *ff; cudaMalloc((void **)&ff, 100 * sizeof(float)); why << <1, 1 >> > (dd, ff); cudaDeviceSynchronize(); return 0; }
So I've simplified everything down to just NSight running locally on one machine, with one 1060 running the minimal code below. With command line cuda-memcheck, all's fine. The only issues I have is when the VS inbuilt NSight (running locally) has memory checker enabled. Also if I NSight to a remote machine, the behaviour remains.

No memory fails too, as you can see.


__global__ void why(
double *dd,
float *ff
)
{
float thingFloat = 3.0f;
float qf = atomicAdd(ff, thingFloat);

double thingDouble = 3.0f;
atomicAdd(dd, thingDouble);
}

int main()
{
double *dd;
cudaMalloc((void **)&dd, 100 * sizeof(double));
float *ff;
cudaMalloc((void **)&ff, 100 * sizeof(float));

why << <1, 1 >> > (dd, ff);

cudaDeviceSynchronize();

return 0;
}

#5
Posted 07/24/2017 05:39 AM   
OK, I get it, it only happens on remote debugging on 1060, right?
OK, I get it, it only happens on remote debugging on 1060, right?

#6
Posted 07/24/2017 06:23 AM   
Ive got it happening locally all on one system now, with a 1060 installed (alongside an old quadro k2000). But if I point it off remotely to a similar system, it still fails. So not the card at least.
Ive got it happening locally all on one system now, with a 1060 installed (alongside an old quadro k2000).
But if I point it off remotely to a similar system, it still fails. So not the card at least.

#7
Posted 07/24/2017 06:29 AM   
Confirmed, bug has been raised.
Confirmed, bug has been raised.

#8
Posted 07/25/2017 08:57 AM   
No repo in our internal build, should be fixed in nsight 5.6
No repo in our internal build, should be fixed in nsight 5.6

#9
Posted 11/28/2017 09:39 AM   
Scroll To Top

Add Reply