macOS: Add required onnxruntime dependency.

[TurboGit]

No description provided.

[zisoft]

please add it to packaging/macosx/1_install_hb_dependencies.sh as well

[TurboGit]

Just done. Thanks!

[kmilos]

The same is needed for Windows? MSYS2 has it, and DLL should be included automatically during packaging...

Scratch that, it's not the runtime.

[andriiryzhkov]

@TurboGit: brew onnxruntime package does not come with CoreML acceleration included. That limits ONNX to running only on CPU on macOS. That would not be a problem for object masking feature, because it does not use acceleration (translation computational graphs to CoreMl costs more time than CPU inference itself), but it might matter for longer tasks like AI denoise.

That was my deliberate decision at some point to get rid of onnxruntime brew package and use binary from GitHub, which comes with CoreML pre-built.

[zisoft]

@andriiryzhkov :

That was my deliberate decision at some point to get rid of onnxruntime brew package and use binary from GitHub, which comes with CoreML pre-built.

How is this supposed to work when darktable is running from the bundle darktable.app?

[andriiryzhkov]

@zisoft :
cmake downloads package from GitHub. But I did not check yet bundle packaging. Looking at it now.

[kmilos]

Sorry for the late feedback here.

IMHO USE_AI should be OFF by default, or at least it shouldn't willy-nilly go downloading stuff without user's explicit request. There's quite a bit of cross-platform packaging stuff to address here.

[zisoft]

The bundling script is packaging/macosx/3_make_hb_darktable_package.sh.
It is all done by simple script commands, no bundling software involved.

[kmilos]

USE_AI should be OFF by default

By this I mean please take it out from .ci/ci-script.sh and put it in ECO where appropriate so it's clearly done with purpose and introduced gradually. For example, we might not want it in all the jobs in the matrix.

option(ONNXRUNTIME_OFFLINE "Disable automatic download of ONNX Runtime" OFF)

Maybe it's just me, but I really don't like the idea of someone else deciding what to download (and bundle) behind my back OOTB, so I'd prefer this to be a user explicit action.

[andriiryzhkov]

It's only package installation during build process. It is either downloaded from brew server or from GitHub. Nothing more.

Event with USE_AI=ON all AI features will be OFF by default in global parameters. To start using AI features "user explicit action" is required to enable AI in parameters and press at least "download default".

[zisoft]

In packaging/macosx/3_make_hb_darktable_package.sh:

./build/_deps/onnxruntime.<version>.dylib needs to be manually copied to the bundle directory.
Somewhere before line 305

[TurboGit]

IMHO USE_AI should be OFF by default

We want the code to at least be compiled during CI builds.

[andriiryzhkov]

@zisoft :

./build/_deps/onnxruntime..dylib needs to be manually copied to the bundle directory.
Somewhere before line 305

working on this

[TurboGit]

@TurboGit: brew onnxruntime package does not come with CoreML acceleration included. That limits ONNX to running only on CPU on macOS. That would not be a problem for object masking feature, because it does not use acceleration (translation computational graphs to CoreMl costs more time than CPU inference itself), but it might matter for longer tasks like AI denoise.

So we will revisit/fix that at some point but the immediate need is to have a fully working binaries.

[kmilos]

We want the code to at least be compiled during CI builds.

Yes, but not even all of the CI builds. Please move to ECO as suggested.

[zisoft]

@TurboGit :

but the immediate need is to have a fully working binaries.

So this PR should fix this until we have a decision

[andriiryzhkov]

I am already testing packaging solution. Will be ready today

[zisoft]

@andriiryzhkov :

brew onnxruntime package does not come with CoreML acceleration included.

Is there a special reason for that?
Maybe a PR at the homebrew package may be the better solution.
I have done that in the past for several other packages (i.e. GMIC)

[andriiryzhkov]

Is there a special reason for that?

The Homebrew onnxruntime formula doesn't include CoreML support – there's an open issue requesting it ([homebrew-core#177367](Homebrew/homebrew-core#177367)) that was closed as "not planned." The flag --use_coreml has to be explicitly passed at build time, and the Homebrew maintainers decided the additional Xcode SDK dependency wasn't worth taking on for a niche feature.

[zisoft]

another way is to rebuild the package from source during the homebrew package install.
I have implemented this with graphicsmagick to make it bundle-able:

darktable/.github/workflows/nightly.yml

Lines 384 to 390 in 1f3fe7a

	# rebuild graphicsmagick for macOS bundling
	export HOMEBREW_NO_INSTALL_FROM_API=1
	gm_formula=`brew edit --print-path graphicsmagick`
	cat $gm_formula | sed 's/--with-modules/--disable-installed/1' > gm_formula.tmp
	mv gm_formula.tmp $gm_formula
	brew reinstall --build-from-source --force graphicsmagick

And this is done for the nightly builds only, not for CI.

[andriiryzhkov]

#20507 fixes ONNX Runtime packaging on macOS

[TurboGit]

@andriiryzhkov : So IIUC I can close this PR in favor of #20507?

[zisoft]

@TurboGit : I think this can be closed now.