Feature request: "Build minmal julia" makefile option

[ldgirod]

It would be nice to have a target in the makefile that built only the base components of julia required to get to the REPL loop with a minimal set of builtins and dependencies.

This came up because I wanted to find out what the minimal requirements are to port Julia to a new architecture (e.g. arm) or platform (e.g. android).

[JeffBezanson]

This is a good idea, but "minimal" means something different to everybody. To some people julia is useless without FFTs, and to others they are bloat that is never used. I would welcome suggestions of which pieces are especially troublesome.

[ldgirod]

This came up in the context of figuring out what's required to port to arm

• so minimal referred to getting the JIT and REPL loop etc. built.

On Thu, May 16, 2013 at 1:29 PM, Jeff Bezanson notifications@github.comwrote:

This is a good idea, but "minimal" means something different to everybody.
To some people julia is useless without FFTs, and to others they are bloat
that is never used. I would welcome suggestions of which pieces are
especially troublesome.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/3128#issuecomment-18016071
.

[JeffBezanson]

This will probably require some significant reorganization of the standard library (which might be worthwhile anyway). We will still have several external dependencies, however, since a lot of the functionality is used internally rather than just exported to the user. This includes PCRE (regular expressions), libm, libunwind, double-conversion, and libuv (for all our I/O). If those are ok, then this might be doable. But eliminating all of those would require very major rewrites.

[ViralBShah]

I think of that as the minimal set too. It seems that this would also not depend on any gpl libs.

[lindahua]

The most important library that Julia relies on is probably BLAS & LAPACK. Even in a "minimal version", they are indispensable.

Here are some efforts to port these to ARM: http://www.vesperix.com/arm/

However, as far as I can tell, they are far less mature than the counterparts for X86 architecture, and LAPACK is still not there.

[lindahua]

I would say at this point the easiest way to use Julia on devices (e.g. smartphones and pads) is to develop a thin client that connects to a web-REPL, instead of porting Julia altogether.

[ldgirod]

In my case, I am doing signal processing on a phone and cannot send all the
raw data to a server; I was looking at Julia because it looks very pleasant
to use and also high performance. At the moment though it looks like it is
non-trivial to port.

On Mon, May 20, 2013 at 9:51 AM, Dahua Lin notifications@github.com wrote:

I would say at this point the easiest way to use Julia on devices (e.g.
smartphones and pads) is to develop a thin client that connects to a
web-REPL, instead of porting Julia altogether.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/3128#issuecomment-18148217
.

[ldgirod]

This guy got LINPACK running on android:

http://specificimpulses.blogspot.com/2011/01/my-android-speaks-fortran-yours-can-too.html

On Mon, May 20, 2013 at 9:47 AM, Dahua Lin notifications@github.com wrote:

The most important library that Julia relies on is probably BLAS & LAPACK.
Even in a "minimal version", they are indispensable.

Here are some efforts to port these to ARM: http://www.vesperix.com/arm/

However, as far as I can tell, they are far less mature than the
counterparts for X86 architecture, and LAPACK is still not there.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/3128#issuecomment-18148034
.

[ViralBShah]

We will probably do all yhe linear algebra using trivial implementations for bigfloat, and that will also serve as a fallback in absence of lapack.

[vtjnash]

closing in favor of #5155

[timholy]

For folks working on Julia's core, it might be interesting to be able to build a really minimal julia...not something any user would want, but effectively a gdb-like mode for Core.

[timholy]

With reference to the above comment, for anyone hacking on inference.jl: just go to the base/ directory and then do include("coreimg.jl"). It will replace the current Core.Inference with a new one.