function Atomics

What kind of problem i have when my GPU support Compute Capability only 1.0 ? I think Atomics function , what is it ?

Yours Sincerely

Rafał Bator

The major limits of Compute 1.0 are very strict global memory coalescing rules, no atomics, and fewer registers per SM.
The programming guide has an appendix which describes all the Compute capability differences.

The major limits of Compute 1.0 are very strict global memory coalescing rules, no atomics, and fewer registers per SM.
The programming guide has an appendix which describes all the Compute capability differences.

what does “coalescing” mean ?

Yours Sincerely

Rafał Bator

what does “coalescing” mean ?

Yours Sincerely

Rafał Bator

When threads from one warp (group of threads) try to access memory, GPU tries to access memory in as short time as possible.

If thread 0 accesses memory at address N+0, thread 1 from N+1, etc., this memory is coalesced and such access is fast.

In case of devices with CC 1.0 this is the only possible speed up. Other devices can have fast access to memory with different patterns of addresses, e.g. with permutations of addresses in the warp, and so on. Fpr details - read CUDA Best Practices.

When threads from one warp (group of threads) try to access memory, GPU tries to access memory in as short time as possible.

If thread 0 accesses memory at address N+0, thread 1 from N+1, etc., this memory is coalesced and such access is fast.

In case of devices with CC 1.0 this is the only possible speed up. Other devices can have fast access to memory with different patterns of addresses, e.g. with permutations of addresses in the warp, and so on. Fpr details - read CUDA Best Practices.