-
Notifications
You must be signed in to change notification settings - Fork 197
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Reduce compile times of distance specializations #1307
Reduce compile times of distance specializations #1307
Conversation
The calculation of the tile indices are now performed in ldgXY(). This will make it possible to remove all state related to the tile index out of the class in the next commit. Note that the calculation of the tile index can depend on which overloaded constructor is called(!)
This commit moves all grid and tile indexing logic into the caller. Contractions_NT is now only responsible for *intra*-tile indexing. Due to the complexity of the epilog function, the ldgNextGridStride function is not yet called from within the main loop. That is the next goal so that we have all the grid and tile indexing localized in the loop.
This commit removes the epilog function and moves its functionality into the run loop. The next step might be to see if the ldgNextGridStride() method has to be called the current location, or if performance is the same if its called at the start of the loop.
This results in subtle issues with non-square KernelPolicy, as found in fusedL2KNN.
This is more general than just for L1. Making use of it more is work in progress.
By adding yet another struct ^^
This did remove support for the CUTLASS kernels. Has to be put back.
I wasted a lot of time because I had not replaced the op::core() method of the l2_exp_distance_op after I copied it from l2_unexp_distance_op... If I copy something from the template and forget to fill it in, I get a compile error.
I am testing on CUDA 12, where it does not seem to work. Prior to my commits, the CUTLASS kernels were also not working. So not sure what's up. In any case: consider this untested.
This indicates that the operator uses expensive operations (pow, exp, log) in the inner loop. Therefore, unrolling and or veclen parameters should be adjusted
No problem! The merge was quite painless actually. Really looking forward to the GTC keynote and all the great progress we will make going forward! |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Two minor comments that I'll leave to you to decide if you wish to implement! LGTM
…he suggestions from dask team
…ild-2304-pin_dask
I think we have to wait for #1363 to clear so that the CI passes for this run. |
@ahendriksen yep, and CI hasn't been particularly friendly lately.... Jobs have been getting queued up for several hours at a time. If it helps, you can merge that branch into this one, then at least this one will be up to date and pass through CI so we can merge it right after. |
…ializations-reduce-compile-times
We should be able to remove the need for the ops-codeowners review once the other PR gets merged |
/merge |
This should cut compilation time for refine_d_int64_t_float.cu.o et al from ~900 seconds to 29 seconds. The refine specialization contain >100 instances of the ivfflat_interleaved_scan kernel, even though these should be seperately compiled by the ivfflat_search specializations. The call to ivf_flat_interleaved_scan is [here](https://github.com/rapidsai/raft/blob/56ac43ad93a319a61073dce1b3b937f6f13ade63/cpp/include/raft/neighbors/detail/refine.cuh#L121). Depends on (so please merge after) PR #1307. Authors: - Allard Hendriksen (https://github.com/ahendriksen) - Corey J. Nolet (https://github.com/cjnolet) Approvers: - Corey J. Nolet (https://github.com/cjnolet) URL: #1360
Following the findings in https://github.com/ahendriksen/raft/tree/investigate-compile-time-reduction-strategies#investigation-of-compile-times, this PR reduces the compile times of the pairwise distance specializations.
This is achieved by:
spdlog
andrmm
are avoided.Additional improvements geared towards iterative development:
raft/distance/distance.cuh
.