question on atomic operation

Accelerated Computing CUDA CUDA Programming and Performance
#1

I’ve been debugging my program for quite a while and finally I’ve found that the problem arises from the code below:

//Eventually counter_device_global gets the total number of new data generated by each thread

atomicAdd(&counter_device_global, newDataSize);

//Write the new data to the array

for(int i=counter_device_global-1;i<counter_device_global-newDataSize;i--)

{

     arrayInPageLockedMemory[i]=newData;

}

The catch is: different threads will read the same counter_device_global, and therefore write to the same memory address. Maybe using counter_device_global is really not a good idea at all, but is there any way that can write data to arrayInPageLockedMemory in a sequential manner? Thanks a lot for any suggestion! :biggrin:

Finally worked out:

//Eventually counter_device_global gets the total number of new data generated by each thread

int temp=atomicAdd(&counter_device_global, newDataSize);

//Write the new data to the array

for(int i=0;i<newDataSize;i++)

{

     arrayInPageLockedMemory[temp+i]=newData;

}
#2

I’d expect you want something like this:

//Eventually counter_device_global gets the total number of new data generated by each thread

int pos=atomicAdd(&counter_device_global, newDataSize);

//Write the new data to the array

for(int i=0;i<newDataSize;i++)

{

     arrayInPageLockedMemory[pos+i]=newData[i];

}

Note that this code will store the data from each thread consecutively, but the order of the data from different threads is undefined.

#3

Hi tera,

Thanks for pointing that out. :thumbup: I copy-pasted the wrong code. Now it’s correct.