Coalesced accesses on different arrays

Accelerated Computing CUDA CUDA Programming and Performance
1

Say that I have two different arrays that will be sent to the device:

__global__ void kernel1(int *array1, int *array2)

And I access each array’s elements the following way:

int a1 = array1[threadIdx.x];

int a2 = array2[threadIdx.x];

Are the accesses being coalesced in a compute capability 1.1 device? I believe so, as each Kth thread accesses the Kth element and there’s is no misalignment while accessing each array…

Is my assumption correct?

2

Yes, assuming you have a 1D block (or, most of the ways of doing non-1D blocks as well!)

3

It will be indeed coalesced. However you should probably see better performance if you make it a int2 array instead of 2 single array.

Take a look here, at MrAnderson tests: http://forums.nvidia.com/index.php?showtop…mp;#entry290441

eyal