-
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
IVF-FLAT support k > 256 #2169
IVF-FLAT support k > 256 #2169
Conversation
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.
Thanks Malte for the PR! It looks great overall, here are my comments
cpp/include/raft/neighbors/detail/ivf_flat_interleaved_scan-inl.cuh
Outdated
Show resolved
Hide resolved
@tfeher thanks for the review. I have added your suggestions. |
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.
Thanks Malte for the updates, it is great to see the increasing code reuse between IVF methods! The PR looks good to me.
As discussed offline, the perf of the non-fused top-k search is currently less then ideal due to large number of dummy values (as a result from cluster size imbalance). We discussed how to improve that in a follow up PR for radix top-k.
/merge |
This PR is a followup to #2169. To enable IVF-flat with k>256 we need an additional select_k invocation which was unexpectedly slow. There are two reasons for that: First problem is the data handed to select_k: The valid data length per row is much smaller than the conservative maximum that could be achieved by probing the N largest probes. Therefore each query row contains roughly ~50% dummy values. This is also the case for IVF-PQ, but did not show up as prominent due to the second reason. The second problem, and also a difference to the IVF-PQ algorithm - is that a 64bit payload data type is used for selectK. The performance of selectK with 64bit index type is significantly slower than with 32bit, especially when many elements are in the same range: ``` Benchmark Time CPU Iterations ----------------------------------------------------------------------------------------------------- SelectK/float/uint32_t/kRadix11bitsExtraPass/1/manual_time 1.68 ms 1.74 ms 413 1357#200000#512 SelectK/float/uint32_t/kRadix11bitsExtraPass/3/manual_time 2.31 ms 2.37 ms 302 1357#200000#512#same-leading-bits SelectK/float/int64_t/kRadix11bitsExtraPass/1/manual_time 5.92 ms 5.98 ms 116 1357#200000#512 SelectK/float/int64_t/kRadix11bitsExtraPass/3/manual_time 83.7 ms 83.8 ms 8 1357#200000#512#same-leading-bits ----------------------------------------------------------------------------------------------------- ``` The data distribution within a IVF-flat benchmark resulted in a select_k time of ~24ms. ### scope: * additional parameter added to select_k to optionally pass individual row lengths for every batch entry. This parameter is utilized by both IVF-Flat and IVF-PQ and results in a ~2x speedup (50 nodes out of 5000) of the final `select_k`. * refactor ivf-flat search to work with 32bit indices by storing positions instead of actual indices. This allows to utilize 32bit index type select_k for ~10x speedup in the final `select_k`. FYI @tfeher @achirkin ### not in scope: * General optimization of select_k: In the current implementation there is no difference in the type of the payload and the actual index type. Especially the type of the histogram has a large effect on performance (due to the atomics). Authors: - Malte Förster (https://github.com/mfoerste4) Approvers: - Tamas Bela Feher (https://github.com/tfeher) URL: #2221
Add support for topk > 256 for ivf_flat (Issue #1555)
The PR adds a non-fused version of topk that is utilized if k > 256.
FYI, @tfeher