-
Notifications
You must be signed in to change notification settings - Fork 21
Cooperative vector API #141
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
@merlinND I forgot to create a separate PR for this part, done now. |
0122307
to
9532b48
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Review part 1 (missing llvm/cuda_coop_vec.cpp
).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I had a look at the two remaining files in this PR. I don't know much about LLVM IR so I mostly reviewed the CUDA backend.
It's really nice how a few helpers make the long intrinsic calls in cuda_coop_vec.cpp
quite compact and readable, despite all the different ops and cases.
Dr.Jit can elide scatter operations when their result can no longer be referenced by any other operations. The logic to do so, and when reference count decreases are needed, was dispersed throughout ``eval.cpp``. This commit simplifies the underlying code.
Dr.Jit previously chose the lowest possible PTX version for each compute capability, but this ended up being too restrictive. It now ships a table containing a full driver version -> PTX version mapping and then searches it for the highest possible PTX version.
This commit adds cooperative vector code generation support to Dr.Jit-Core. The main documentation and test suite of this feature is in the Dr.Jit parent project.
- Use continuation callables when compiling OptiX programs that use the cooperative vector extension. This yields slightly better performance. - When the program uses wide (>64) CoopVec networks, set a high register target to avoid spills.
This PR adds cooperative vector code generation support to Dr.Jit-Core. The main documentation and test suite of this feature is in the Dr.Jit parent project.