On Thu, Dec 31, 2015 at 2:38 PM, Oded Gabbay <oded.gab...@gmail.com> wrote:
> On Thu, Dec 31, 2015 at 11:30 AM, Oded Gabbay <oded.gab...@gmail.com> wrote:
>> On Wed, Dec 30, 2015 at 5:41 PM, Roland Scheidegger <srol...@vmware.com>
>> wrote:
>>> Am 30.12.2015 um 10:46 schrieb Oded Gabbay:
>>>> On Wed, Dec 30, 2015 at 1:11 AM, Roland Scheidegger <srol...@vmware.com>
>>>> wrote:
>>>>>
>>>>> So, if I see that right, you will automatically generate binaries using
>>>>> power8 instructions if compiled on power8 capable box, which then won't
>>>>> run on boxes not supporting power8? Is that really what you want?
>>>>> Maybe some runtime detection would be a good idea (though I don't know
>>>>> if anyone cares about power7)?
>>>>
>>>> The problem is I don't think I can eliminate the build time check
>>>> (although I would very much like to) because I need:
>>>> 1. To pass a special flag to the GCC compiler: -mpower8-vector
>>>> 2. To define _ARCH_PWR8 so GCC will include the newer intrinsic
>>>>
>>>> Without those two things, I won't be able to use vec_vgbbd which I
>>>> need to implement the _mm_movemask_epi8 efficiently, and without that,
>>>> all this patch series can be thrown out the window. The emulation of
>>>> _mm_movemask_epi8 using regular instructions is just horrible.
>>>>
>>>> You are correct that once you build a binary with this flag on power8
>>>> machine, that binary won't run on power7 machine. You get "cannot
>>>> execute binary file"
>>>>
>>>> Unfortunately, I don't see a way around this because even if I
>>>> condition the use of vec_vgbbd on a runtime check/define, the library
>>>> still won't be executable because it was built with -mpower8-vector.
>>>>
>>>> Having said that, because I *assume* IBM right now mostly cares about
>>>> Linux running on POWER8 with little-endian, I think it is a fair
>>>> compromise.
>>>
>>> Note I don't have anything against a build time check. My concern here
>>> is something along the lines of unsuspecting distros shipping binaries
>>> which won't work, as it looks to me like this will get picked up
>>> automatically. That is different to how for instance sse41 is handled.
>>> That is I believe this should only get enabled if someone has specified
>>> some -mcpu=power8 or whatever flag explicitly somewhere already.
>>>
>>> Roland
>>
>> I understand and I share your concern. Maybe we should add
>> "--disable-pwr8-inst" to mesa's configure ? if that flag is given to
>> configure, it would disable the optimization code (won't add
>> _ARCH_PWR8 to defines and won't add -mpower8-vector to gcc flags).
>>
>> What do you think ?
>>
>> Oded
>
> Actually, I made a mistake in checking this issue. I forgot my power8
> machine is LE and power7 is BE - that's why the binary couldn't be
> executed.
>
> I need to install a power8 BE and re-check this. If the binary can be
> executed, than I will add runtime checks.
>
> Oded
>
So indeed the problem was LE vs. BE. Once I built it on power8 BE, I
could run it on power7 (which is always be). Of course, it crashed
with illegal instruction, but I can fix that using runtime detection.
I will send a new patch series with all the fixes.
Oded
>>
>>>
>>>>
>>>> Oded
>>>>
>>>>> So far we didn't bother with that for SSE
>>>>> but it has to be said SSE2 is a really low bar (and the manual assembly
>>>>> stuff doesn't use anything more advanced, even though clearly things
>>>>> like the emulated mm_mullo_epi32 are suboptimal if your cpu supports
>>>>> sse41). And even then on non-x86 you actually might not get
>>>>> PIPE_ARCH_SSE if you didn't set gcc's compile flags accordingly.
>>>>>
>>>>> Roland
>>>>>
>>>>>
>>>>> Am 29.12.2015 um 17:12 schrieb Oded Gabbay:
>>>>>> To determine if we could use special POWER8 assembly directives, we first
>>>>>> need to detect whether we are running on POWER8 architecture. This patch
>>>>>> adds this detection to configure.ac and adds the necessary compilation
>>>>>> flags accordingly.
>>>>>>
>>>>>> Signed-off-by: Oded Gabbay <oded.gab...@gmail.com>
>>>>>> ---
>>>>>> configure.ac | 30 ++++++++++++++++++++++++++++++
>>>>>> 1 file changed, 30 insertions(+)
>>>>>>
>>>>>> diff --git a/configure.ac b/configure.ac
>>>>>> index f8a70be..1acd47e 100644
>>>>>> --- a/configure.ac
>>>>>> +++ b/configure.ac
>>>>>> @@ -396,6 +396,36 @@ fi
>>>>>> AM_CONDITIONAL([SSE41_SUPPORTED], [test x$SSE41_SUPPORTED = x1])
>>>>>> AC_SUBST([SSE41_CFLAGS], $SSE41_CFLAGS)
>>>>>>
>>>>>> +dnl Check for POWER8 Architecture
>>>>>> +PWR8_CFLAGS="-mpower8-vector"
>>>>>> +have_pwr8_intrinsics=no
>>>>>> +AC_MSG_CHECKING(whether we are running on POWER8 Architecture)
>>>>>> +save_CFLAGS=$CFLAGS
>>>>>> +CFLAGS="$PWR8_CFLAGS $CFLAGS"
>>>>>> +AC_COMPILE_IFELSE([AC_LANG_SOURCE([[
>>>>>> +#if defined(__GNUC__) && (__GNUC__ < 4 || (__GNUC__ == 4 &&
>>>>>> __GNUC_MINOR__ < 8))
>>>>>> +#error "Need GCC >= 4.8 for sane POWER8 support"
>>>>>> +#endif
>>>>>> +#include <altivec.h>
>>>>>> +int main () {
>>>>>> + vector unsigned char r;
>>>>>> + vector unsigned int v = vec_splat_u32 (1);
>>>>>> + r = __builtin_vec_vgbbd ((vector unsigned char) v);
>>>>>> + return 0;
>>>>>> +}]])], have_pwr8_intrinsics=yes)
>>>>>> +CFLAGS=$save_CFLAGS
>>>>>> +
>>>>>> +if test $have_pwr8_intrinsics = yes ; then
>>>>>> + DEFINES="$DEFINES -D_ARCH_PWR8"
>>>>>> + CXXFLAGS="$CXXFLAGS $PWR8_CFLAGS"
>>>>>> + CFLAGS="$CFLAGS $PWR8_CFLAGS"
>>>>>> +else
>>>>>> + PWR8_CFLAGS=
>>>>>> +fi
>>>>>> +
>>>>>> +AC_MSG_RESULT($have_pwr8_intrinsics)
>>>>>> +AC_SUBST([PWR8_CFLAGS], $PWR8_CFLAGS)
>>>>>> +
>>>>>> dnl Can't have static and shared libraries, default to static if user
>>>>>> dnl explicitly requested. If both disabled, set to static since shared
>>>>>> dnl was explicitly requested.
>>>>>>
>>>>>
>>>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
