Skip to content
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.

Sign up for GitHub

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

Jump to bottom

CAGRA binary size too large. #1459

Open
Tracked by #1392
tfeher opened this issue Apr 24, 2023 · 1 comment
Open
Tracked by #1392

CAGRA binary size too large. #1459

tfeher opened this issue Apr 24, 2023 · 1 comment
Assignees
@enp1s0
Labels
Vector Search

Comments

@tfeher
Copy link
Contributor

tfeher commented Apr 24, 2023 •
edited
Loading

Note that rapids is compiled for sm 60, 70, 75, 80, 86, 90. Compiling for float only, with one team specilization per team adds 125 MiB to libraft.so. Four data types and multiple team size specialization goes over 1 GiB size increase. Find out what is essential for good perf, and what is acceptable size increase for RAFT.
@tfeher tfeher mentioned this issue Apr 24, 2023
Open
16 tasks
@tfeher tfeher changed the title Binary size too large. Note that rapids is compiled for sm 60, 70, 75, 80, 86, 90. Compiling for float only, with one team specilization per team adds 125 MiB to libraft.so. Four data types and multiple team size specialization goes over 1 GiB size increase. Find out what is essential for good perf, and what is acceptable size increase for RAFT. CAGRA binary size too large. Apr 24, 2023
@tfeher tfeher mentioned this issue Apr 25, 2023
Closed
@tfeher
Copy link
Contributor Author

tfeher commented Apr 25, 2023

See #1428 (comment) details in binary size (when compiled only for float input type).

If we could have block size as runtime param (instead of template) that would lead a significant reduction in binary size
https://github.com/rapidsai/raft/blob/branch-23.06/cpp/include/raft/neighbors/detail/cagra/search_single_cta.cuh#L827-L841

@tfeher tfeher assigned tfeher and enp1s0 and unassigned tfeher Apr 25, 2023
@achirkin achirkin mentioned this issue May 10, 2023
Merged
@tfeher tfeher mentioned this issue May 10, 2023
Merged
rapids-bot bot pushed a commit that referenced this issue Jun 9, 2023
@tfeher
CAGRA pad dataset for 128bit vectorized load (#1505)
6ec78e9 
This PR adds padding to the dataset (if necessary) to make reading any of its rows compatible with 128bit vectorized loads. This change also enables handling arbitrary number of input features (before this PR each row had to be at least 64bit aligned, which constrained the acceptable number of input features).

Fixes #1458.

With this change, it is sufficient to keep a single "load type" specialization for the search kernels, which shall cut the binary size by half (#1459).

Authors:
  - Tamas Bela Feher (https://github.com/tfeher)

Approvers:
  - tsuki (https://github.com/enp1s0)
  - Corey J. Nolet (https://github.com/cjnolet)

URL: #1505
@cjnolet cjnolet mentioned this issue Jun 22, 2023
Closed
9 tasks
@tfeher tfeher mentioned this issue Jul 17, 2023
Merged
@enp1s0 enp1s0 mentioned this issue Aug 16, 2023
Merged
rapids-bot bot pushed a commit that referenced this issue Sep 26, 2023
@enp1s0
Remove block size template parameter from CAGRA search (#1740)
6c7cada 
This PR removes block size template parameters from CAGRA search kernel functions to reduce the library size and build time.

rel: #1459

Authors:
  - tsuki (https://github.com/enp1s0)

Approvers:
  - Tamas Bela Feher (https://github.com/tfeher)

URL: #1740
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@enp1s0 enp1s0

Labels
Vector Search
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@cjnolet @tfeher @enp1s0