mlir/lib/Dialect/GPU/Transforms: improve context management in SerializeToCubin #65779
Conversation
There was a problem hiding this comment.
The patch itself LGTM, however, we are on the process of deprecating SerializeToCubin in favor of Target Attributes. I'm introducing today deprecation notices. So I'm -1 on improving the existing passes, in my opinion all new efforts should focus on the new mechanism. However, don't know if @joker-eph or someone else has a different opinion.
|
The patch is small enough that it seems worthwhile to take in, I would just want to make sure we don't diverge from the lowering done through the new flow: do we need to replicate this somewhere as well @fabianmcg ? |
|
Currently no, as we don't invoke the driver. However, I was thinking on adding a compilation path to stop at PTX and let the driver JIT the code at runtime, I only need to do some small updates, so maybe then. |
That's fine -- i'm currently using it (as part of the reference pipeline), so I am incentivized to make it faster. I don't understand the CI failure, can I get some help with that? |
|
It's an infra failure, feel free to ignore |
|
Actually your PR is not rebased, seems like you're based on a commit from May! |
rohany
force-pushed
the
serialize-cubin-context-management
branch
from
91bcd17 to
fb0a003
Compare
…izeToCubin This commit adjusts the CUDA context management in the SerializeToCubin pass. In particular, it uses the device 0 primary context instead of creating a new CUDA context on each invocation of SerializeToCubin. This yields very large improvements in compile time, especially if an application (like a JIT compiler) is calling SerializeToCubin repeatedly. Differential Revision: https://reviews.llvm.org/D159487
rohany
force-pushed
the
serialize-cubin-context-management
branch
from
fb0a003 to
5e1a41b
Compare
|
Thanks, fixed it. |
|
@rohany Do you need help to commit this change? |
|
Yes, i don't know how to get it to land, given that tests pass + accepted review. |
Local branch amd-gfx 319c66a Merged main:080fb3e5b73b into amd-gfx:7c4daea7af99 Remote branch main 71bdd2c mlir/lib/Dialect/GPU/Transforms: improve context management in SerializeToCubin (llvm#65779)
This commit adjusts the CUDA context management in the SerializeToCubin pass. In particular, it uses the device 0 primary context instead of creating a new CUDA context on each invocation of SerializeToCubin. This yields very large improvements in compile time, especially if an application (like a JIT compiler) is calling SerializeToCubin repeatedly.
Differential Revision: https://reviews.llvm.org/D159487