-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
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
Stop hiding math intrinsics from llvm #19640
Labels
Comments
simonbyrne
added
maths
Mathematical functions
broadcast
Applying a function over a collection
labels
Dec 18, 2016
Yes, it's pretty much #12830 + patches to LLVM to tell it which libm to use. |
Is this a 1.0 issue? |
See also JuliaMath/Libm.jl#8 |
mbauman
added
compiler:simd
instruction-level vectorization
and removed
broadcast
Applying a function over a collection
labels
Apr 24, 2018
We don't have as many intrinsics anymore, as many have now been translated into Julia. Additionally, we now use Cassette for this ("should find a proper solution ... and switch back"). Closing in favor of #15265. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Currently we hide math intrinsics (e.g.
exp
,log
, etc) from LLVM to avoid it using the wrong libm for constant folding and in the JIT. However, esp on highly parallel architectures such as KNL and probably on GPUs as well, LLVM can vectorize those functions and in fact needs to if we want to get performance. However, this of course doesn't work if LLVM doesn't know that those functions exist. We should find a proper solution to it using the wrong libm and switch back to using the intrinsics.The text was updated successfully, but these errors were encountered: