[Mesa-stable] [Mesa-dev] [PATCH] swr: fix -march flag for AVX

[Rowley, Timothy O]

> On Jun 12, 2016, at 3:56 AM, Steven Newbury <steve at snewbury.org.uk> wrote:
> 
> On Fri, 2016-06-10 at 16:05 -0400, Ilia Mirkin wrote:
>> On Fri, Jun 10, 2016 at 3:43 PM, Tim Rowley <timothy.o.rowley at intel.c
>> om> wrote:
>>> 
>>> Previously used core-avx-i was for ivybridge;
>>> corei7-avx allows sandybridge.
>>> ---
>>>  src/gallium/drivers/swr/Makefile.am | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>> 
>>> diff --git a/src/gallium/drivers/swr/Makefile.am
>>> b/src/gallium/drivers/swr/Makefile.am
>>> index d211f2e..8156cf2 100644
>>> --- a/src/gallium/drivers/swr/Makefile.am
>>> +++ b/src/gallium/drivers/swr/Makefile.am
>>> @@ -124,7 +124,7 @@ COMMON_LDFLAGS = \
>>>  lib_LTLIBRARIES = libswrAVX.la libswrAVX2.la
>>> 
>>>  libswrAVX_la_CXXFLAGS = \
>>> -       -march=core-avx-i \
>>> +       -march=corei7-avx \
>> Just wondering if it'd be enough to say like
>> 
>> -march=x86_64
>> -mavx
>> 
>> and add -mavx2 for libswrAVX2.
>> 
>> I suspect you've iterated through this 20 times and this has some
>> shortcoming I'm not thinking of, but figured I'd point it out just in
>> case it helps.
>> 
>>   -ilia
>> 
> Maybe I'm the only one who finds it horrible to override -march from
> within project build systems.  It causes no end of problems with LTO,
> and results in objects being built inappropriately for the target as
> specified by the builder.

I agree that the way swr is built can be improved and welcome ideas/discussion for work on the master branch, but for the moment I’m trying to get the 12.0 branch in good shape for swr with non-invasive changes (as we’ve tested the current setup pretty extensively).

> Why not use function attributes?  Either compile a function
> specifically for a particular target, or specifically enable AVX though
> "__attribute__ ((__target__ ("avx")))" (ideally with a fall-back
> implementation detected at run-time when the instruction set isn't
> supported).  Even better, have it also protected at build-time by a
> simple check for __AVX__ so the code can be compiled out entirely where
> the package builder has specified -march(=native) as a build flag.

One problem with the function attribute path is that as far as I know, MSVC does not have a similar feature.

-Tim

>>> 
>>>         -DKNOB_ARCH=KNOB_ARCH_AVX \
>>>         $(COMMON_CXXFLAGS)