Global CMake for tutorials

[cedricchevalier19]

The goal is to provide a joint built for all exercises (and their solutions).

This way, compilation is straightforward for IDE users.

• Currently, most of the exercises are built using the top-level CMake.

• However, not all builds make sense, as CUDA builds are wrong for the first exercises. I have put configuration warnings when Kokkos is not configured correctly, as explaining why it is not working is still valuable.

• I have also refactored the fetch_content implementation: it is now called in an overloaded find_package to keep CMakeLists.txt as standard as possible.

Three CMake variables can be set by the user to control the build system:

• KokkosTutorials_FORCE_INTERNAL_Kokkos: to force a build of a custom Kokkos and ignore any installed Kokkos
• KokkosTutorials_FORCE_EXTERNAL_Kokkos: to force finding an installed Kokkos and to fail if not found
• KokkosTutorials_KOKKOS_SOURCE_PATH: to specify a Kokkos source directory and by-pass fetch_content when building a custom Kokkos.

I can improve the global CMake by using external_project for the first exercises to "sandbox" their builds and allow a custom Kokkos.

[cedricchevalier19]

One difficulty is that the first 3-4 exercises are not perfect Kokkos code and can have issue when a Kokkos device backend is enabled.

[cedricchevalier19]

I will add Kokkos-Kernels tutorials before merge, but I want to know other thoughts before continuing.

A lot of files are touched, but their modifications are pretty similar.

[masterleinad]

Why not the latest release? Did you intend to update in a separate pull request?

[cedricchevalier19]

Yes, I thought a separate PR made more sense.

[dalg24]

How common do we expect it is that users will have their IDE setup to (remotely?) target a GPU for the tutorials?

[dalg24]

Same comment here, please propose this change as its own PR.

[cedricchevalier19]

This change was necessary in order to compile this exercise. I will comment out the add_subdirectory and do a separate PR.

[cedricchevalier19]

How common do we expect it is that users will have their IDE setup to (remotely?) target a GPU for the tutorials?

I cannot put any number, but laptops with decent GPUs are getting more popular, so I think compiling and trying the exercises on the local computer is natural.

[ajpowelsnl]

Cédric - a few general comments before diving into an 89-file review:

• Would it make sense to radically narrow this PR, and convert one of the exercises to CMake to make your proposed changes easier to understand? We just need one strong proof-of-concept example to understand how we could / should extend to the other exercises
• With the one CMake-i-fied exercise example, demonstrate correct linking and runtime behavior for the target backend
• It's likely our shortcoming that we don't specify in the tutorials how students should build Kokkos. Would it make sense to point students to the Quick Start at the start of the tutorial for a semi-standard Kokkos configure / build / install

[cedricchevalier19]

• Would it make sense to radically narrow this PR, and convert one of the exercises to CMake to make your proposed changes easier to understand? We just need one strong proof-of-concept example to understand how we could / should extend to the other exercises

It would have. However, I do not think it is this useful to do it now that @dalg24 and @masterleinad have spent to much time browsing many files (sorry ...).

• It's likely our shortcoming that we don't specify in the tutorials how students should build Kokkos. Would it make sense to point students to the Quick Start at the start of the tutorial for a semi-standard Kokkos configure / build / install

That's a great idea. I will add it in the README, but in another PR.

[ajpowelsnl]

• Would it make sense to radically narrow this PR, and convert one of the exercises to CMake to make your proposed changes easier to understand? We just need one strong proof-of-concept example to understand how we could / should extend to the other exercises

It would have. However, I do not think it is this useful to do it now that @dalg24 and @masterleinad have spent to much time browsing many files (sorry ...).

• Let us not succumb to the "sunk costs" fallacy 😄 ; the fact that others have begun reviewing the unwieldy PR is not a real argument against doing one simple, clear, correct example. Would you consider closing this PR, and using what you learned here to create one proof-of-concept example?

• It's likely our shortcoming that we don't specify in the tutorials how students should build Kokkos. Would it make sense to point students to the Quick Start at the start of the tutorial for a semi-standard Kokkos configure / build / install

That's a great idea. I will add it in the README, but in another PR.

Excellent. The more we can standardize and "automate" the tutorial/education process (e.g., even in ChatGPT) the greater the success we'll have.

[dalg24]

• Would it make sense to radically narrow this PR, and convert one of the exercises to CMake to make your proposed changes easier to understand? We just need one strong proof-of-concept example to understand how we could / should extend to the other exercises

It would have. However, I do not think it is this useful to do it now that @dalg24 and @masterleinad have spent to much time browsing many files (sorry ...).

• Let us not succumb to the "sunk costs" fallacy 😄 ; the fact that others have begun reviewing the unwieldy PR is not a real argument against doing one simple, clear, correct example. Would you consider closing this PR, and using what you learned here to create one proof-of-concept example?

It does not make sense to split it and only tackle one example, regardless of how much time we spent on this before.

[ajpowelsnl]

• Would it make sense to radically narrow this PR, and convert one of the exercises to CMake to make your proposed changes easier to understand? We just need one strong proof-of-concept example to understand how we could / should extend to the other exercises

It would have. However, I do not think it is this useful to do it now that @dalg24 and @masterleinad have spent to much time browsing many files (sorry ...).

• Let us not succumb to the "sunk costs" fallacy 😄 ; the fact that others have begun reviewing the unwieldy PR is not a real argument against doing one simple, clear, correct example. Would you consider closing this PR, and using what you learned here to create one proof-of-concept example?

It does not make sense to split it and only tackle one example, regardless of how much time we spent on this before.

• Separating the question of time from the discussion, if each exercise is stand alone, I don't understand why a number of moderately correct / incorrect examples would be more instructive (as a proof-of-concept) than one clear, more correct example

[dalg24]

• Would it make sense to radically narrow this PR, and convert one of the exercises to CMake to make your proposed changes easier to understand? We just need one strong proof-of-concept example to understand how we could / should extend to the other exercises

It would have. However, I do not think it is this useful to do it now that @dalg24 and @masterleinad have spent to much time browsing many files (sorry ...).

• Let us not succumb to the "sunk costs" fallacy 😄 ; the fact that others have begun reviewing the unwieldy PR is not a real argument against doing one simple, clear, correct example. Would you consider closing this PR, and using what you learned here to create one proof-of-concept example?

It does not make sense to split it and only tackle one example, regardless of how much time we spent on this before.

• Separating the question of time from the discussion, if each exercise is stand alone, I don't understand why a number of moderately correct / incorrect examples would be more instructive (as a proof-of-concept) than one clear, more correct example

This PR is about providing a parent project that includes all tutorials, and applying to all of them is more or less applying the same changes multiple places. Feel free to speak up if you don't like the general direction of this PR, but, so you know, I would strongly object to not tackling all (or least most) in this PR and I read Cedric's polite objection as him agreeing with me on that matter.

[JBludau]

can this be moved into a separate function instead of repeating it?

[dalg24]

Like a kokkos_tutorials_warn_cpu_only_program() function or whatnot?

[cedricchevalier19]

The idea is that each exercise can be used as a standalone. I think factorizing with a function or macro makes the example less clear, and I wish I could avoid including a separate (but shared) file at the terminal CMakeLists.txt.

[dalg24]

It some way it is akin to the weird find kokkos module isn't it? I'd slightly prefer reusing code as Jakob suggested

[cedricchevalier19]

I had trouble finding the proper solution as the Kokkos_DEVICES variable is not apparent. If I do a function, I will hide the details that might be useful to copy and paste for some users. That's the reason why it is better to be transparent there.
There are only two such exercises; we will probably not have to touch the CMakeLists.txt too often, so the function is not worth it.

However, I can add a comment explaining what this conditional does for better logic readability (which is the function's main advantage for me).

[dalg24]

Like a kokkos_tutorials_warn_cpu_only_program() function or whatnot?

[dalg24]

just out of curiosity why "dot"cmake for the directory name?

[cedricchevalier19]

I am notoriously bad at naming things. Initially, I wanted to put the helper files in a classic cmake directory, but having such a utility directory in the middle of the exercise was confusing. So, I thought it was clever to hide it with a starting dot.
Any better naming scheme is welcomed.

[dalg24]

I'd prefer w/o the dot

[JBludau]

How about just renaming the targets 01_Exercise -> 01_Begin and 01_Solution, but leaving the build system as it is. For IDE users we can still provide a IDE/CMakeLists.txt file with something like:

cmake_minimum_required(VERSION 3.16)
project(KokkosTutorialAll)

add_subdirectory(../Exercises/01/Begin)
add_subdirectory(../Exercises/01/Solution)
...

The warning message for the device backends could go into the common.cmake.

But this still requires us to adapt the slides for the tutorials and the only gain is if someone wants to build all exercises at the same time.