Symptoms of bank conflicts in Shared Memory

Accelerated Computing CUDA CUDA Programming and Performance
1

Is there any signs / symptoms if bank conflicts occur in the program?
eg. Speed slower than non-shared memory version

I want to get some hints before learning and installing another toolkit.

Thanks a lot.

2

Hi, good question.

I searched a little bit but didn’t find the exact latency if bank conflicts occur.

So perhaps I’m going to write a benchmark for that.

If on simply does a vector copy from A to B and goes over shared memory, a strided access of 16 should hit one bank all the time, thus pretty bad.

However I don’t think you will get below global memory.

But let’s try it.

:-)