I have implemented FCC in CUDA and wanted to use shared memory to enhance the speed up but due to overlapping memory access by threads I have bank conflicts. Is there a way I can use texture memory to enhance the speed up. In essence I want to know does texture memory support simultaneous memory accesses by threads or it also serializes them just like shared memory does?
Sachin Gupta
I don’t think you can getter better performance using Textures if you are able to keep your data in shared memory and playing with it without the need of writing/ reading to/from global memory. Also, if your bank conflicts are not of a higher degree, you should keep it this way, else try for another approach that helps you reduce the bank conflicts.
I don’t think you can getter better performance using Textures if you are able to keep your data in shared memory and playing with it without the need of writing/ reading to/from global memory. Also, if your bank conflicts are not of a higher degree, you should keep it this way, else try for another approach that helps you reduce the bank conflicts.
Thanks a lot for your comments.
Can you help me out to know what order of bank conflicts we are facing if 1 thread access a 1-d array from say 1-16 index and the other thread is accessing from 2-17.In my particular case potential solution lies in further lowering down the box size but is the above degree of bank conflict too high?
I know that if all the threads work in total synchronization then each one will access a position adjacent to each other which lie in different banks but in actual conditions such a synchronization is not possible so what could be the predicted order of bank conflict
Thanks in advance