Some apps are missing CMake builds

[alexreinking]

The following apps do not have CMake builds and so are not properly tested on CI

• HelloAndroid (not built by Makefile either)
• HelloAndroidCamera2 (not built by Makefile either)
• HelloAndroidGL (not built by Makefile either)
• HelloMatlab
• HelloPyTorch (not built by Makefile either)
• HelloiOS (not built by Makefile either)
• HelloWasm (has CMake build, not built by Makefile)
• auto_viz (not built by Makefile either)
• fft
• hexagon_benchmarks (not built by Makefile either)
• hexagon_dma (not built by Makefile either)
• linear_blur (not built by Makefile either)
• nn_ops (not built by Makefile either)
• onnx
• opengl_demo (not built by Makefile either)
• openglcompute (not built by Makefile either)
• resnet_50
• simd_op_check (not built by Makefile either)

There might be good reasons for some of these, eg. I think linear_blur uses generator stubs, which aren't available from CMake (yet/ever). But even then, we should document that somewhere.

[abadams]

Note that for HelloMatlab the makefile only exists as a stub to get the buildbots to call run_blur.sh. Matlab/octave itself takes over the build. A cmakelists with a single custom target that calls the script should suffice?

HelloiOS is intended as an example xcode project, but I'm not convinced it's doing things in a good way - it bakes in a shell script that builds the generator as a custom build step. I'm not sure if cmake can generate full ios applications natively. xcode keeps changing things like where the toolchain lives. HelloAndroid is somewhat similar - it's an attempt to use the android native build tools (at the time, gradle). I fear it has rotted.

simd_op_check appears to be a minimal stub to get an AOT version of test/correctness/simd_op_check.cpp to run on an android device using the object files output by that test.

I think those are the weird ones. Everything else should be more straightforward.

[steven-johnson]

I fear it has rotted.

I will bet you $1 that all the iOS and Android-specific builds are totally rotten. Is there a way to target these via CMake?

[abadams]

I did recently use HelloiOS as a stub for an iOS app with success, and then spent several hours obsessively tweaking the reaction diffusion equation involved to do different pretty things. I think I may have had to adjust a path somewhere, but apart from that it was working. Getting it cmake'd instead of custom shell scripts would be great if it's possible and maintainable.

[abadams]

(Note that we can't have cmake take over the whole project build though - the point of the app is how to incorporate Halide into an existing xcode project)

[alexreinking]

Note that for HelloMatlab the makefile only exists as a stub to get the buildbots to call run_blur.sh. Matlab/octave itself takes over the build. A cmakelists with a single custom target that calls the script should suffice?

Yes, I think that should suffice. Will have to add an equivalent script for Windows, though.

HelloiOS is intended as an example xcode project, but I'm not convinced it's doing things in a good way - it bakes in a shell script that builds the generator as a custom build step. I'm not sure if cmake can generate full ios applications natively. xcode keeps changing things like where the toolchain lives. HelloAndroid is somewhat similar - it's an attempt to use the android native build tools (at the time, gradle). I fear it has rotted.

I know that CMake can generate Xcode projects and first-class support for iOS, watchOS, and tvOS. So that should be workable.

simd_op_check appears to be a minimal stub to get an AOT version of test/correctness/simd_op_check.cpp to run on an android device using the object files output by that test.

CMake also has first-class Android support.

[alexreinking]

I will bet you $1 that all the iOS and Android-specific builds are totally rotten. Is there a way to target these via CMake?

Yes. CMake natively supports both iOS and Android.

(Note that we can't have cmake take over the whole project build though - the point of the app is how to incorporate Halide into an existing xcode project)

One could use CMake to generate an Xcode project and then include that one as a subproject? I admit I have no iOS development experience, so that might not be a great workflow.

[abadams]

I'd be happy to try out that xcode configuration, but I'm also not an iOS development person. I do have an ipad though that I can try things on.

[yangyanli]

Note that for HelloMatlab the makefile only exists as a stub to get the buildbots to call run_blur.sh. Matlab/octave itself takes over the build. A cmakelists with a single custom target that calls the script should suffice?

HelloiOS is intended as an example xcode project, but I'm not convinced it's doing things in a good way - it bakes in a shell script that builds the generator as a custom build step. I'm not sure if cmake can generate full ios applications natively. xcode keeps changing things like where the toolchain lives. HelloAndroid is somewhat similar - it's an attempt to use the android native build tools (at the time, gradle). I fear it has rotted.

simd_op_check appears to be a minimal stub to get an AOT version of test/correctness/simd_op_check.cpp to run on an android device using the object files output by that test.

I think those are the weird ones. Everything else should be more straightforward.

HelloiOS failed to compile in my case at:
gpu_tile(clobber.x, clobber.y, 1, 1)
with error:
xxx/Halide/apps/HelloiOS/reaction_diffusion_2_generator.cpp:135:18: no matching member function for call to 'gpu_tile'

[yangyanli]

Note that for HelloMatlab the makefile only exists as a stub to get the buildbots to call run_blur.sh. Matlab/octave itself takes over the build. A cmakelists with a single custom target that calls the script should suffice?

HelloiOS is intended as an example xcode project, but I'm not convinced it's doing things in a good way - it bakes in a shell script that builds the generator as a custom build step. I'm not sure if cmake can generate full ios applications natively. xcode keeps changing things like where the toolchain lives. HelloAndroid is somewhat similar - it's an attempt to use the android native build tools (at the time, gradle). I fear it has rotted.

simd_op_check appears to be a minimal stub to get an AOT version of test/correctness/simd_op_check.cpp to run on an android device using the object files output by that test.

I think those are the weird ones. Everything else should be more straightforward.

also, I have to add:
-isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX11.1.sdk \
to the Run Script for resolving an issue similar to this.