Fixed pernicious bug in autobatching #1127
Merged
Conversation
This file contains 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
|
Does this actually affect things? If these are executed on the same stream,
the kernel shouldn’t start executing until after the copy completes. Is the
story that this forced synch is expensive?
…On Fri, 8 Dec 2017 at 4:56 am, Graham Neubig ***@***.***> wrote:
OK, it took me literally all day to find a bug that can be fixed in 10
characters of code.
Basically, parallel memcpy (used in autobatching) requires memory to be
transferred to the GPU, but this was being done in an asynchronous fashion,
and as a result the memory wasn't necessarily arriving at the GPU before
the memcpy kernel started executing.
This probably fixes #1057 <#1057>
------------------------------
You can view, comment on, or merge this pull request online at:
#1127
Commit Summary
- Fixed pernicious bug in autobatching
File Changes
- *M* dynet/exec.cc
<https://github.com/clab/dynet/pull/1127/files#diff-0> (16)
Patch Links:
- https://github.com/clab/dynet/pull/1127.patch
- https://github.com/clab/dynet/pull/1127.diff
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#1127>, or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAJba4ntwNqRtJ40Bb--2BJjnYWVu10jks5s-MGFgaJpZM4Q6lxk>
.
|
|
I'm not actually sure whether the copy and the kernel are being executed on the same stream (probably not). Probably the best thing to do is enforce that the copy uses the same stream as the kernel, but I wasn't immediately sure how to do this. And yes, I'm aware that the forced sync might be expensive, I haven't merged because I haven't had time to figure out the speed ramifications yet. |
|
OK, so CudaMemcpy caused a performance degradation, but specifying a stream for CudaMemcpyAsync seems to be working for now. I'm not super-confident in this fix though, so after merging I'll continue to monitor for reported problems. |
Closed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
OK, it took me literally all day to find a bug that can be fixed in 10 characters of code.
Basically, parallel memcpy (used in autobatching) requires memory to be transferred to the GPU, but this was being done in an asynchronous fashion, and as a result the memory wasn't necessarily arriving at the GPU before the memcpy kernel started executing.
This probably fixes #1057