Repository navigation
Conversation
a10y
force-pushed
the
aduffy/patched-fuse
branch
from
April 9, 2026 19:22
e16d1de to
721fa3d
Compare
a10y
marked this pull request as ready for review
April 10, 2026 16:52
a10y
added
changelog/feature
0ax1
reviewed
Apr 13, 2026
a10y
force-pushed
the
aduffy/patched-fuse
branch
2 times, most recently
from
April 13, 2026 20:58
3acf500 to
88567cf
Compare
0ax1
approved these changes
Apr 14, 2026
0ax1
left a comment
0ax1
left a comment
Contributor
There was a problem hiding this comment.
LGTM - Added a couple more thoughts.
| /// This kernel uses a thread-per-lane model where each thread is assigned to | ||
| /// one (chunk, lane) slot and applies all patches in that slot. | ||
| template <typename ValueT> | ||
| __device__ void patched(ValueT *const output, |
Contributor
There was a problem hiding this comment.
Nice! Ideally we could move to a world where we drop the other patches.cu eventually and only operate on transposed patches on the GPU, maybe.
| match_each_integer_ptype!(bitpacked.ptype(bitpacked.dtype()), |P| { | ||
| return decode_bitpacked::<P>( | ||
| bitpacked.into_owned(), | ||
| P::default(), |
Contributor
There was a problem hiding this comment.
We could also handle the fused case here right: FoR + BP?
| } | ||
|
|
||
| // Execute the components | ||
| let lane_offsets = array |
Contributor
There was a problem hiding this comment.
This could go in a separate fn:
let lane_offsets = array
.lane_offsets()
.clone()
.execute_cuda(ctx)
.await?
.into_primitive()
.into_data_parts()
.buffer;
let patch_indices = array
.patch_indices()
.clone()
.execute_cuda(ctx)
.await?
.into_primitive()
.into_data_parts()
.buffer;
let patch_values = array
.patch_values()
.clone()
.execute_cuda(ctx)
.await?
.into_primitive()
.into_data_parts()
.buffer;
a10y
force-pushed
the
aduffy/patched-fuse
branch
from
April 17, 2026 13:47
3462e80 to
f6f1ccf
Compare
Signed-off-by: Andrew Duffy <andrew@a10y.dev>
a10y
force-pushed
the
aduffy/patched-fuse
branch
from
April 17, 2026 13:48
f6f1ccf to
120e84a
Compare
Contributor
Author
|
Benchmark results for 
|
2 tasks done
Contributor
Author
|
I think we're going to change layout of Patched array to be sorted w/chunk offsets instead of lane-split. See #7440 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This fixes up CUDA execution to adaptively support interior or exterior patches for bit unpacking.
This does not implement Dynamic Dispatch