Improve CUDA softmax performance #4973
Merged
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
apaszke
force-pushed
the
softmax_speedup
branch
from
db2e612 to
80a5fc4
Compare
|
All the build failures are spurious and unrelated to this PR |
colesbury
approved these changes
Feb 1, 2018
|
|
||
| inline dim3 SoftMax_getBlockSize(int ILP, uint64_t dim_size) { | ||
| uint64_t block_size = 1; | ||
| uint64_t max_block_size = std::min(dim_size / ILP, static_cast<uint64_t>(1024)); |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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Outdated suggestions cannot be applied.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.
Simple fix with a very large perf benefit for smaller sizes. Below are some plots (
dim_size= size of the softmaxed dimension,outer_size= batch size, z-axis = ratio of old time to new time). In general, as long asdim_size < 1024you get at least a 2x speedup with this code, 4x if you fit in 256, and even 12x for sizes around 100 and smaller.I tried playing with some other potential improvements like replacing the blockReduce function with a shuffle-based one, but it gave mixed results (-20% time in some cases, +20% time in other cases).
Thanks to @nikitakit for reporting #4893 (which is fixed in this PR).