Register usage for vector types

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

Do I understand correctly register consumption for the following declarations inside kernels:

ushort2 var1;            consumes ONE  32-bit register?

unsigned short var2[2];  consumes TWO  32-bit registers?

If above definitions were in shared/global memory, either of them would consume total 32 bits of memory…

Thank you

2

You might be looking at compiler optimization in the first case. If you don’t use all the elements of a vector type, the compiler can and will optimize the unused elements away. This can lead to strange side effects, including confusing register usage and uncoalesced memory access.