[RFC] How to add vector math functions to Glibc

[Andrew Senkevich]

2014-11-06 23:51 GMT+03:00 Andrew Senkevich <andrew.n.senkevich@gmail.com>:
> Hi, Joseph,
>
> 2014-10-24 1:37 GMT+04:00 Joseph S. Myers <joseph@codesourcery.com>:
>
>> On Thu, 23 Oct 2014, Andrew Senkevich wrote:
>
>>> diff --git a/math/bits/mathcalls.h b/math/bits/mathcalls.h
>>> index 8a94a7e..2d31a11 100644
>>> --- a/math/bits/mathcalls.h
>>> +++ b/math/bits/mathcalls.h
>>> @@ -60,6 +60,15 @@ __MATHCALL (atan,, (_Mdouble_ __x));
>>>  __MATHCALL (atan2,, (_Mdouble_ __y, _Mdouble_ __x));
>>>
>>>  /* Cosine of X.  */
>>> +#if !defined _Mfloat_ && !defined _Mlong_double_ && defined __DECL_SIMD_cos
>>> +__DECL_SIMD_cos
>>> +#endif
>>> +#if defined _Mfloat_ && !defined _Mlong_double_ && defined __DECL_SIMD_cosf
>>> +__DECL_SIMD_cosf
>>> +#endif
>>> +#if defined _Mlong_double_ && defined __DECL_SIMD_cosl
>>> +__DECL_SIMD_cosl
>>> +#endif
>>>  __MATHCALL (cos,, (_Mdouble_ __x));
>>
>> As previously noted, I think it would be much better if the definition of
>> __MATHCALL can include all the conditional bits (possibly through a
>> generated header that defines __DECL_SIMD_cos etc. to empty if not defined
>> by bits/math-vector.h).
>
> proposal is to use separated __MATHCALL_VEC for vector cases, because
> it reduces number of needed empty definitions and can be simply
> generated (__MATHCALL case requires a lot of manual search to obtain
> all affected function names because of redefinitions in some files).
>
>>> +#if defined __x86_64__ && defined __FAST_MATH__
>>> +# if defined _OPENMP && _OPENMP >= 201307
>>> +/* OpenMP case. */
>>> +#  define __DECL_SIMD_AVX2 _Pragma("omp declare simd notinbranch")
>>> +#  define __DECL_SIMD_SSE4 _Pragma("omp declare simd notinbranch")
>>
>> Of course we still need the API/ABI documentation providing the stable
>> guarantee about exactly what this pragma means regarding the function
>> versions it is saying are available in glibc.
>
> We will follow-up on this soon.
>
> I attached patch with almost all infrastructure fixes discussed
> before. It seems pragma meaning and data tables remain to be done.
> Patch affects a lot of files and of course will be separated to
> minimal disjoint parts for submission later.

Here is the patch updated in part of data table and function code
accordingly points mentioned before in this discussion.


--
WBR,
Andrew
-------------- next part --------------
A non-text attachment was scrubbed...
Name: libmvec_141114.patch
Type: application/octet-stream
Size: 52589 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20141114/933e6540/attachment.obj>