Is it worth it to unwind a static loop with a lot of branches?
I have a static loop with a lot of branches inside. If I unwind it, and flatten out obvious branches, it becomes sequential code. Is branch prediction good enough to avoid following every single path in each iteration of the loop or do I have to do that manually by unwinding the loop?
In general, the more sequential the code (no branching) the better the performance. Remember that on the CPU chip, L1 cache is generally used for instructions - and the size of such is limited. L2 cache can also be used, but the goal is to fit it into the L1 for best performance. With the branch situation you've described, within the loop, all the possible branches will be loaded - which will take up much more room than the unwound loop.
I've not used CUDA (I've done some reading), so I cannot claim that the above will apply, but my understanding is that it's just function calls passing in all the necessary parameters for the work to be done. According to the chart (under "Version features and specifications"), http://en.wikipedia.org/wiki/CUDA, the cache is 8K.
Branch prediction does not even exists on GPUs. The GPU thread scheduler will just switch execution to a different warp until the outcome of the branch has been resolved.
However unrooling loops can still be beneficial:
- Unrolling can reduce the loop overhead by removing unnessecary compare and conditonal branch instructions
- Unrolling might allow the GPU to access more ILP and can result in good performance even with a smaller number of active warps per SM
- Unrolling might allow to place some arrays into registers instead of DRAM: indexed acceses to registers is not possible at the moment and unrolling can sometimes allow the compiler to resolve the indexes at compile time
The L1 code cache is quite small, too aggressive unrolling can also cause performance issues. It can also reduce occupancy, if the unrolled code needs more registers.
But you can easily check what unrolling does for the performance of your code:
You do not need to manually unroll your code, just place "#pragma unroll" in front of your loop if you want a complete unroll of the loop or "#pragma unroll 4" if you want 4 iterations of your loop to be unrolled and so on.
Can be. You can coerse the compiler to unroll a for{} loop for you using "#pragma unroll". You'll find the details of that in the Cuda Programming Guide.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Computer Software