Add alpaka::getPreferredWarpSize(dev)
#2216
Conversation
alpaka::getPreferredWarpSize(dev) returns one of the possible warp sizes supported by the device. On devices that support a single work size (cpu, CUDA gpu, ROCm gpu), getPreferredWarpSize(dev) avoids the overhead of wrapping that value in an std::vector. On devices that support multiple warp sizes, the value returned by getPreferredWarpSize(dev) is unspecified. Currently it returns the largest supported value -- but this could change in a future version of alpaka. Signed-off-by: Andrea Bocci <[email protected]>
fwyzard
force-pushed
the
alpaka_getPreferredWarpSize
branch
from
5e2db10 to
2b368fc
Compare
|
I am just curious on the purpose of this API. Is the main goal to avoid the heap allocation of Because we could just change the API to either return e.g. a |
| @@ -181,10 +182,22 @@ namespace alpaka::trait | |||
| auto find64 = std::find(warp_sizes.begin(), warp_sizes.end(), 64); | |||
| if(find64 != warp_sizes.end()) | |||
| warp_sizes.erase(find64); | |||
| // Sort the warp sizes in decreasing order | |||
| std::sort(warp_sizes.begin(), warp_sizes.end(), std::greater<>{}); | |||
There was a problem hiding this comment.
[Nit] If the vector size can be large somehow, "initially sorting then finding the element in vector" (logarithmic time) is faster compared to "finding deleting then sorting" (linear time). After sorting, using std::lower_bound function finds in log(n) time.
There was a problem hiding this comment.
The largest vector I encountered had 5 elements: { 4, 8, 16, 32, 64 }.
There was a problem hiding this comment.
@mehmetyusufoglu your analysis is correct, but I am with @fwyzard: the supported warp sizes are probably a small set here :) Also, binary search is slower than linear for small sizes due to data dependent access pattern. So @fwyzard's version is probably faster for our use case :D
| auto find64 = std::find(warp_sizes.begin(), warp_sizes.end(), 64); | ||
| if(find64 != warp_sizes.end()) | ||
| warp_sizes.erase(find64); |
There was a problem hiding this comment.
Btw, in C++20, this should be just std::erase(warp_sizes, 64). Looking forward to the upgrade :)
|
@fwyzard if you want the PR merged, please mark the PR as Ready for review, thx! |
|
Thanks for the review. I've marked it as a draft because I want to figure out first how it interacts with caching the device information. |
alpaka::getPreferredWarpSize(dev)returns one of the possible warp sizes supported by the device.On devices that support a single work size (cpu, CUDA gpu, ROCm gpu),
getPreferredWarpSize(dev)avoids the overhead of wrapping that value in anstd::vector.On devices that support multiple warp sizes, the value returned by
getPreferredWarpSize(dev)is unspecified. Currently it returns the largest supported value -- but this could change in a future version of alpaka.Add a test for
alpaka::getPreferredWarpSize(dev).