What does "Allocation Granularity" mean in CUDA Occupancy Calculator


I saw the Allocation Granularity item in CUDA Occupancy Calculator on the tab of GPU Data. It states that compute capability lower than 2.0 would have “Allocation Granularity” of blocks, while 2.0 has that of warps. What does “Allocation Granularity” mean here? The allocation of space? The scheduling scheme of threads? Thanks.

It’s the allocation granularity of registers. 1.x devices round the numbers of registers per block, 2.x devices rounds the number of registers per warp.

Check section 4.2 of the Programming Guide.