Skip to content

[DRAFT] Shared memory buffering of candidates and selected items in DeviceTopK #8326

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

Draft
elstehle wants to merge 9 commits into
base: main
Choose a base branch
from
+924 −200
Draft

[DRAFT] Shared memory buffering of candidates and selected items in DeviceTopK #8326

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
decaca3
extracts histo into own kernel
elstehle Apr 7, 2026
b449e38
[bench-only] benchmarking changes
elstehle Apr 7, 2026
671b105
[bench-only] benchmarking run
elstehle Apr 7, 2026
4cad24d
[bench-only] adds smem buffering before gmem writes
elstehle Apr 8, 2026
597a429
fixes determining keys_only
elstehle Apr 10, 2026
d76869f
introduces new smem mode
elstehle Apr 10, 2026
52546b4
[bench-only] Merge remote-tracking branch 'upstream/main' into enh/to…
elstehle Apr 10, 2026
5eaa259
lower shmem requirements
elstehle Apr 10, 2026
a4ba5fc
[bench-only] no smem-coalescing for 8B offset types
elstehle Apr 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 8 additions & 6 deletions ci/bench.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,10 +29,8 @@ benchmarks:
filters:
# CUB C++ benchmark filters (regex matched against ninja target names).
cub:
# Examples:
# - '^cub\.bench\.for_each\.base'
# - '^cub\.bench\.reduce\.(sum|min)\.'

- '^cub\.bench\.topk\.keys\.base'
- '^cub\.bench\.topk\.pairs\.base'
# Python benchmark filters (regex matched against paths under benchmarks/).
python:
# Examples:
Expand All @@ -47,8 +45,8 @@ benchmarks:
# - "rtxa6000" # sm_86, 48 GB
# - "l4" # sm_89, 24 GB
# - "rtx4090" # sm_89, 24 GB
# - "h100" # sm_90, 80 GB
# - "rtxpro6000" # sm_120
- "h100" # sm_90, 80 GB
- "rtxpro6000" # sm_120

# Extra .devcontainer/launch.sh -d args
# launch_args: "--cuda 13.1 --host gcc14"
Expand All @@ -64,4 +62,8 @@ benchmarks:
--stopping-criterion entropy
--throttle-threshold 90
--throttle-recovery-delay 0.15
--axis "OffsetT{ct}=I32"
--axis "SelectedElements[pow2]=[3:23:5]"
--axis "Entropy=[1.000,0.201,0.000]"
--axis "OutOffsetT{ct}=I32"
nvbench_compare_args: ""
Loading
Loading