Tested with:

• TensorFlow-1.13.1-foss-2019a-Python-3.7.2.eb
• TensorFlow-1.15.2-fosscuda-2019b-Python-3.7.4.eb
• TensorFlow-2.1.0-foss-2019b-Python-3.7.4.eb
• TensorFlow-2.2.0-foss-2019b-Python-3.7.4.eb
• TensorFlow-2.2.0-fosscuda-2019b-Python-3.7.4.eb

Thanks for the testing. This change should currently have no effect as the dependencies are missing. I added them to the 2.1.0 ECs in easybuilders/easybuild-easyconfigs#11109 so you can add those to your tests too if you like.

Edit: While doing this I noticed that there is a compiler wrapper for intel which sets some env vars. The better approach would be to use --action_env which trims down the ~100 lines to about 4. I'm also suspicious about 'intel_license_file': os.getenv('INTEL_LICENSE_FILE', os.getenv('LM_LICENSE_FILE')),. Is it really ok to set INTEL_LICENSE_FILE=$LM_LICENSE_FILE (and removing $LM_LICENSE_FILE) basically renaming the var? That could also be fixed by that change

Also another idea: As TF likes to add/remove/rename packages, maybe we should add all names to the EasyBlock and verify the list. Even when we don't have a corresponding EC yet.

I addressed your comments and also some changes:

• introduce function get_system_libs_for_version which does the version filtering and can be used to render a list that can be diffed against the TF included list on updates. IMO crucial to keep up to date. We might even want to do this automatically during the build, what do you think?
• Add ALL TF known libs to a) allow for the above and b) see which are missing to potentially add new ECs or find the correct ones
• Combined the 2 lists into 1 and use values of None and True to denote missing ECs and "assume installed"
• Change order of tf_name and eb_name and sort lines. There are benefits in doing it this way (e.g. check against the tf list) but maybe the other way round is better as it allows to see name changes more easily as EC names stay (usually) constant (see jpeg). I tend to revert the order again but I'm still not 100% happy with the list anyway. Not great to read. Any ideas?

Ok, still need to do a few adjustments. I'm adding and testing more ECs to be used as dependencies. To check I added a log which dependencies that could be used by TF are not.

The main dependency I'd like to include is protobuf but that needs another patch and a test. Checking now if that works...

@boegel needed to rebase to include the ninja fix to test this. New commits are starting at "Address review"

Doing a last round of testing, but this should be good to go now... Thanks a lot for your efforts on this @Flamefire!

This has been tested extensively, both with a wide range of existing TensorFlow easyconfigs and the updated ones for TensorFlow 2.1.0 from easybuilders/easybuild-easyconfigs#11109 that require these changes, so going in!