[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Can a leaf package require SSE2 on i386?



On 09/15/2014 04:04 PM, Mathieu Malaterre wrote:
> On Sun, Sep 14, 2014 at 8:47 AM, Sébastien Villemot
> <sebastien@debian.org> wrote:
>> Hi,
>>
>> As the maintainer of julia (a technical computing language built on top
>> of LLVM), I am wondering whether I should continue supporting the i386
>> architecture.
>>
>> The bottom line is that julia needs SSE2 (and porting it to the x87 FPU
>> requires changes that are beyond what I am willing/able to do, see [1]
>> for more details). And the presence of SSE2 is not guaranteed on the
>> i386 architecture.
> 
> As per ld.so man page:
> 
> HARDWARE CAPABILITIES
>        Libraries might be compiled using hardware-specific
> instructions which do not exist on all CPU. Such libraries should be
> installed in directories whose name defines
>        the hardware capabilities such as /usr/lib/sse2/. The dynamic
> linker checks these directories against the hardware of the machine
> and selects the best suitable ver‐
>        sion of a given library. Hardware capabilities directories
> could be cascaded to combine CPU features. Hardware capabilities
> depends on the CPU. The following  names
>        are currently recognized:
> [...]
>       x86 (32-bit only)
>               acpi, apic, clflush, cmov, cx8, dts, fxsr, ht, i386,
> i486, i586, i686, mca, mmx, mtrr, pat, pbe, pge, pn, pse36, sep, ss,
> sse, sse2, tm
> 
> HTH

Thanks. That's nice, however, what if a library needs SSE4.2? Two of my
packages, eg libjerasure2 and libgf-complete1, would benefit from major
speed-up if there was SSE4 available. I have currently disabled all SSE
stuff, and would really like to provide an SSE4.2 version as well. I
suppose (according to what's above) that using
/usr/lib/sse4.2/x86_64-linux-gnu isn't supported (yet), right? (note: I
already asked upstream if it was possible to do runtime detection, and
the answer is currently no, unfortunately)

Cheers,

Thomas Goirand (zigo)


Reply to: