Thanks!

I would be in favour of merging progressively the patches.
@jzstark @ryanrhymes do you have a strong opinion on this?

@kkirstein can you update the appropriate opam files to add a lower bound on both the new conf-openblas and "eigen" {>= 0.3.0}?

Fantastic, thanks

I am going to merge these changes. I will make a new release once windows passes all the tests

Thanks to Kay-Uwe for investigating and helping with this issue. Looks like the C macros I have implemented a long time ago needs some improvement on the Windows OS.

However, I find that with the recent code, several examples such as mnist_cnn.ml generate error: Exception: Failure "input_value: unknown custom block identifier". I don't think the problem is introduced in this PR. Perhaps it's just my machine/version mismatch/bad network connection etc., but it does require some further investigation. (Also IMO the INSTALL_WINDOWS.md should be combined into the INSTALL.md file, but that's just trivial matter.)

I will try to track down this problem when I'm available. But be it a bug or not, it does involve the issue of proper code reviewing. Perhaps it goes unnoticed by some of us, but recently Liang and I have just added a set of project governance rules, involving the team structure, PR, making releases, etc. E.g. one of them requires that a PR be reviewed by at least two team members. (I was planning to check this PR tonight, and thought we can at least wait for a while on it :)

Currently we are considering restructuring the team, since the working environment has changed a lot for both of us for now. And this governance is part of this management procedure. I do believe that it will be followed by the community from now on, so as to ensure the team can achieve a mutual understanding of the current project progress, and that this project can grow healthily in a controllable manner.

I have identified the mentioned crash and added as issue #550 so we can track it.
@jzstark: Also for me it looks like there could be an issue with involved C macros, probably, there are different default definitions in mingwcompared to "vanilla" gcc.