Thanks, that solves my problem! With your solution, I find most of the
execution time is spent in from the body of the C function, not calling the
function, so I'm not giving up much by not inlining it.
On Wednesday, February 19, 2014 1:05:40 PM UTC-8, Isaiah wrote:
>
> Hi Bob,
>
> If you only need the digamma function, you could do this:
>
> ```
> // fastbigoneheader.c, compile with `gcc -fPIC -shared fastbigoneheader.c
> -o fboh.so`
> #include "fastbigoneheader.h"
> float fboh_digamma(float x) { return fastdigamma(x); }
> ```
>
> and then in Julia:
> ```
> fastgamma(x) = ccall( (:fboh_digamma, "fboh.so"), Float32, (Float32,), x)
> ```
>
> But if you are doing this in a very tight loop there is a possibility that
> the function execution could cost less than the calling overhead. So
> porting this function to Julia might result in faster execution overall if
> it ends up inlined.
>
> (for other functions this may not be true, as I noticed that library uses
> __m128 SIMD types in some places, which cannot yet be used from Julia).
>
> Best,
> Isaiah
>
>
> On Wed, Feb 19, 2014 at 3:41 PM, Bob Quazar <[email protected]<javascript:>
> > wrote:
>
>> Hi Ivar,
>>
>> Thanks for the reply. I'm interested in that particular header file
>> because it contains a fast implementation of the "digamma" function. Julia
>> has a built-in digamma function, but I suspect it's orders of magnitude
>> slower than the one implemented in this header file---Julia's digamma is
>> probably much more accurate than I need. (My program's run time is
>> currently dominate by the calls to the digamma function, and there's no
>> easy way to call digamma less often, such as through caching, etc. I'm
>> implementing an optimization procedure for learning parameters a Bayesian
>> model, using stochastic variational inference.)
>>
>> Indeed, I'm already using a version of Julia that I've compiled. I'm not
>> sure yet whether I'll ultimately distribute my code---it would be nice to
>> have the option to do so though. It sounds like it might be easier for me
>> to port the fast digamma function from C to Julia, though I'm a bit
>> disappointed it isn't more straightforward to call this C code from Julia.
>> If there's another way, I'd love to hear it!
>>
>> Bob
>>
>>
>>
>> On Wednesday, February 19, 2014 11:49:50 AM UTC-8, Ivar Nesje wrote:
>>>
>>> I do not think you can use that C code (as is) from Julia. All the
>>> functions are declared "static inline", and thus they need to be compiled
>>> together with the code that calls them.
>>>
>>> I do not have a deep knowledge of C, linking, and shared libraries, but
>>> I will attempt to give an outline on what steps you will need to google for.
>>>
>>> 1. Make the functions externally callable. This means that you will
>>> at least remove `static`, but likely you want to remove `inline also`.
>>> Just
>>> search for `static inline ` and replace by empty string
>>> 2. Compile the c code and link it into the julia process. If you are
>>> on windows you might need to specially mark the symbols as exported. If
>>> you
>>> compile julia yourself, and do not plan to distribute the code, you
>>> might
>>> have success with just #include <fastonebigheader.h>, but ideally you
>>> should compile it as a shared library.
>>> 3. Use the functions with `ccall` as normal. If you include the
>>> header in a julia, you will not need to specify a library.
>>>
>>> You ask a fairly open question that could actually be the basis of a
>>> book. It would be easier to answer more specifically if you included a
>>> paragraph about your project and why you want to use this code. If it is
>>> not just curiosity about whether it is possible, you might instead get a
>>> better suggestion for how to solve your problem.
>>>
>>> Ivar
>>>
>>> kl. 19:22:56 UTC+1 onsdag 19. februar 2014 skrev Bob Quazar følgende:
>>>>
>>>> I'm trying to call C code from Julia. The C code is just a big header
>>>> file: http://fastapprox.googlecode.com/svn/trunk/fastapprox/src/
>>>> fastonebigheader.h
>>>> (There's no corresponding .c file.) I've installed this header file in
>>>> my "/usr/include/" directory, but according to the Julia web site, Julia
>>>> only accesses C shared objects. I'm not sure what to do---wish I
>>>> understood
>>>> C/gcc better. I'd greatly appreciate any help.
>>>>
>>>> Thanks, Bob
>>>>
>>>
>
