Howto use caffeine: documentation and roadmap?

[amckinstry]

Hi,

I've gotten caffeine/coarrays working on Debian Linux with flang-22:

-- llvm-toolchain-22-22.1.1/debian/rules.orig 2026-05-10 10:40:25.428061790 +0100
+++ llvm-toolchain-22-22.1.1/debian/rules 2026-05-10 10:40:42.667293081 +0100
@@ -640,7 +640,6 @@
ifneq (,$(filter $(DEB_HOST_ARCH), $(FLANG_ARCHS)))
PROJECTS += ;flang
RUNTIMES += ;flang-rt

• STAGE_ALL_CMAKE_EXTRA += -DFLANG_ENABLE_PRIF=ON
  endif

alastair@sid:~/deb-packages/llvm$ cat hello.f90
program main
  implicit none
  write (*,*) "Hello from image", this_image(), "of", num_images()
end program main

$ flang-22 -o hello hello.f90 -fcoarray -lcaffeine-smp -lgasnet-smp-seq -L/usr/lib/aarch64-linux-gnu/fortran/flang/ -lassert
warning: Support for multi image Fortran features is still experimental and in development.

$ alastair@sid:~/deb-packages/llvm$ gasnetrun_smp -n 2 ./hello
 Hello from image 1 of 2
 Hello from image 2 of 2 

I've not found any documentation of how its supposed to work; is this it for the moment? is there any work ongoing to simplify this, remove the need to specify libraries etc?

[bonachea]

Hi @amckinstry -
Thanks for you interest!

As stated in the warning message:

warning: Support for multi image Fortran features is still experimental and in development.

The questions you raise are exactly what we mean by "experimental": documentation and scripting support are still lacking. Also the primary features themselves are still under development: flang 22 does not yet include support for lowering coarray variables. As such, it's likely a bit too early yet to be of interest to general end users.

Coming soon!

[amckinstry]

Thanks. I understand its experimental and not yet of general interest. I'd like to help integrate and test in Debian: across architectures, different accelerators, making it easier to use, etc.

A roadmap of where its going and what state it currently is at would be useful. It sounds like the next round of development is in flang/MLIR/LLVM rather than the PLIR library, correct?

[bonachea]

The current state of flang's multi-image support is documented here. Caffeine implements a much broader set of multi-image features, although a few remain to be completed.

Current flang development includes adding lowering support for coarrays and related intrinsics, for example in llvm/llvm-project#182110 and llvm/llvm-project#192944 .

The "roadmap" is basically to incrementally complete implementation of the features listed in the pages linked above, and (once we reach a more usable subset) add sufficient documentation and scripting to make them user-friendly.