[XLA:GPU] Support Numpy-order argsort through CUB via key packing #36043
Conversation
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
copybara-service
bot
force-pushed
the
test_853108960
branch
2 times, most recently
from
c0b0dea to
9b937a6
Compare
copybara-service
bot
changed the title
copybara-service
bot
force-pushed
the
test_853108960
branch
3 times, most recently
from
d560b4e to
b2fd8d0
Compare
FR: #35587 Adds support for CUB-accelerated argsort with Numpy order (NaNs last) for F16, BF16, and F32 keys with S16 or S32 indices. This is implemented by packing the key (converted to an order-preserving unsigned integer) and the index into a single U32 or U64 payload. This allows us to use the standard fast CUB radix sort on the packed pairs. Microbenchmark: ``` Device: NVIDIA_H100_80GB_HBM3 Speedups Clean Dirty name argsort_numpy_order_1024_f32 1.00x 9.8 us 9.8 us argsort_numpy_order_1048576_f64 1.00x 564.7 us 565.1 us argsort_numpy_order_25690112_f64 1.00x 22826.0 us 22835.3 us argsort_numpy_order_1024_f64 1.00x 13.4 us 13.4 us argsort_numpy_order_1024_bf16 1.29x 9.6 us 7.5 us argsort_numpy_order_1024_f16 1.44x 11.1 us 7.7 us argsort_numpy_order_1048576_f32 3.32x 388.9 us 117.2 us argsort_numpy_order_1048576_bf16 5.24x 337.9 us 64.5 us argsort_numpy_order_1048576_f16 5.79x 378.0 us 65.3 us argsort_numpy_order_25690112_f32 8.63x 15299.5 us 1772.2 us argsort_numpy_order_25690112_bf16 12.74x 12155.9 us 954.1 us argsort_numpy_order_25690112_f16 13.67x 13248.3 us 969.0 us ``` PiperOrigin-RevId: 855076701
copybara-service
bot
force-pushed
the
test_853108960
branch
from
b2fd8d0 to
5925e1e
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
[XLA:GPU] Support Numpy-order argsort through CUB via key packing
FR: #35587
Adds support for CUB-accelerated argsort with Numpy order (NaNs last) for F16, BF16, and F32 keys with S16 or S32 indices.
This is implemented by packing the key (converted to an order-preserving unsigned integer) and the index into a single U32 or U64 payload. This allows us to use the standard fast CUB radix sort on the packed pairs.
Microbenchmark: