Re: [flint-devel] Thread safe FLINT

[Michael A. <mab...@go...>]

On Wed, Feb 11, 2009 at 11:39 AM, Bill Hart <goo...@go...> wrote:

Hi Bill,

> I have just committed changes to the development version of FLINT to
> make the test_support and memory_manager code thread safe.

Cool. Can one turn off the use of threads or is this code deeply
interwoven with the rest of the code? I am asking because I am keeping
an eye out for the native Windows port.

> I've verified that all tests in fmpz_poly-test can now run in parallel
> without issues and fixed numerous issues with thread safety in that
> module. All of the functions in fmpz_poly are fully threadsafe, with
> the exception of:
>
> * _fmpz_poly_mul_modular is not threadsafe (it stores static
> information - Gonzalo and I are working on fixing it).
>
> Some of the biggest issues were:
>
> * mpn_random2 is not threadsafe and segfaults
> * the stack based memory manager in FLINT was not threadsafe as it did
> not set up per thread stacks
> * use of randstate_t's in test support without initialising per thread
> was not threadsafe
> * everything needs cleaning up when joining threads (i.e. the per
> thread memory managers and random states), otherwise you get a memory
> leak - there is now a macro in flint.h for doing this cleanup, though
> OpenMP does not seem to have a mechanism for doing automatic cleanup
> on joining threads, so this is the user's responsibility
>
> I have not checked the thread safety of any other part of FLINT,
> however I am unaware of anything else that is not threadsafe except
> the random functions in long_extras and the stack based mpz_t
> allocator in the mpz_extras module (which no one uses anyway).
>
> It is my hope that FLINT 1.2 will be fully threadsafe.
>
> Note gcc 4.2.4 or higher is currently needed to compile the
> development version of FLINT. I'll put some checking in for lower
> versions in FLINT 1.2 before releasing (and OpenMP stuff just won't be
> available to people with older versions of gcc).

Mhh, mandating gcc 4.2 or higher would be a problem for Sage since
currently we only need C99 support, i.e. gcc 3.4.x or higher. If this
dependency comes out of the need for OpenMP and it cannot be turned
off via some config options this is a major issue.

> I haven't decided whether to have the tests to run in parallel or not.
> Currently some of them run much faster in parallel, and some much
> slower.

:)

> Bill.

Cheers,

Michael

> ------------------------------------------------------------------------------
> Create and Deploy Rich Internet Apps outside the browser with Adobe(R)AIR(TM)
> software. With Adobe AIR, Ajax developers can use existing skills and code to
> build responsive, highly engaging applications that combine the power of local
> resources and data with the reach of the web. Download the Adobe AIR SDK and
> Ajax docs to start building applications today-http://p.sf.net/sfu/adobe-com
> _______________________________________________
> flint-devel mailing list
> fli...@li...
> https://lists.sourceforge.net/lists/listinfo/flint-devel
>