Warp shuffle faster than register?

Accelerated Computing CUDA CUDA Programming and Performance
1

I know global memory slower than shared memory, slower than register. but where is warp shuffle? faster than register?

2

Warp shuffle is the same performance as shared memory. Warp shuffle cannot be faster than a register as the input is a read from the register file and the output is a write to the register file.

2 Likes
3

Warp shuffle is the same performance as shared memory.

Typically better. That is, AFAICR it’s faster than a store to shared memory followed by a store from shared memory.

But yes, using registers is best, if one can manage to keep things in registers

1 Like
4

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.