[dpdk-dev] [PATCH] lib/librte_sched: Fix compile with gcc 4.3.4

[Ananyev, Konstantin]

> -----Original Message-----
> From: Qiu, Michael
> Sent: Friday, November 27, 2015 3:22 PM
> To: Ananyev, Konstantin
> Cc: Thomas Monjalon; dev at dpdk.org
> Subject: Re: [PATCH] lib/librte_sched: Fix compile with gcc 4.3.4
> 
> Sorry for not explaining clearly.
> 
> For gcc version start from version 4.4, x86intrin.h will be include, and inside x86intrin.h, immintrin.h will be directly include without
> check AVX

Yes, but inside immintrin.h there is a check.
At least that what I am seeing for gcc4.4.3:

#ifdef __AVX__
#include <avxintrin.h>
#endif

Isn't it the same for your case? 
Konstantin

> (as I know, AVX is not exist when gcc >= 4.4),so no AVX macro does not mean vector disable.
> 
> Only gcc < 4.4 and no macro AVX will disable vector.
> 
> This is my understanding, may be wrong :)
> 
> Thanks,
> Michael
> 
> > 在 2015年11月27日，下午10:09，Ananyev, Konstantin <konstantin.ananyev at intel.com> 写道：
> >
> >
> >
> >> -----Original Message-----
> >> From: Qiu, Michael
> >> Sent: Friday, November 27, 2015 2:02 PM
> >> To: Ananyev, Konstantin
> >> Cc: Thomas Monjalon; dev at dpdk.org
> >> Subject: Re: [PATCH] lib/librte_sched: Fix compile with gcc 4.3.4
> >>
> >> I just replied that Marco AVX only exist in gcc version < 4.4 , otherwise it will not exist.
> >
> > If macro __AVX__ not defined, then
> > #if defined(__AVX__)
> > would always be false and SCHED_VECTOR_ENABLE also wouldn't be defined.
> > So still don't understand why that is a problem
> > Konstantin
> >
> >>
> >> What's your suggest will not work if gcc version greater than 4.3.
> >>
> >> So still need to check gcc version. Any other solution?
> >>
> >> Thanks,
> >> Michael
> >>
> >>> 在 2015年11月27日，下午8:34，Ananyev, Konstantin <konstantin.ananyev at intel.com> 写道：
> >>>
> >>>
> >>>> -----Original Message-----
> >>>> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Qiu, Michael
> >>>> Sent: Friday, November 27, 2015 11:53 AM
> >>>> To: Thomas Monjalon
> >>>> Cc: dev at dpdk.org
> >>>> Subject: Re: [dpdk-dev] [PATCH] lib/librte_sched: Fix compile with gcc 4.3.4
> >>>>
> >>>> really？I don't think so.
> >>>>
> >>>> AVX Marco only exist in the gcc version below 4.4,  I still need to check if below or beyond 4.4 am I right?
> >>>>
> >>>> Thanks,
> >>>> Michael
> >>>
> >>>
> >>> If you look at lib/librte_eal/common/include/arch/x86/rte_vect.h, you'll see the code similar
> >>> to one you are trying to put into rte_shed.c:
> >>>
> >>> lib/librte_eal/common/include/arch/x86/rte_vect.h:
> >>> ...
> >>> #if (defined(__ICC) || (__GNUC__ == 4 &&  __GNUC_MINOR__ < 4))
> >>>
> >>> #ifdef __SSE__
> >>> #include <xmmintrin.h>
> >>> #endif
> >>>
> >>> #ifdef __SSE2__
> >>> #include <emmintrin.h>
> >>> #endif
> >>>
> >>> #ifdef __SSE3__
> >>> #include <tmmintrin.h>
> >>> #endif
> >>>
> >>> #if defined(__SSE4_2__) || defined(__SSE4_1__)
> >>> #include <smmintrin.h>
> >>> #endif
> >>>
> >>> #if defined(__AVX__)
> >>> #include <immintrin.h>
> >>> #endif
> >>>
> >>> #else
> >>>
> >>> #include <x86intrin.h>
> >>>
> >>> #endif
> >>> ...
> >>>
> >>> So I think you can do just like that:
> >>>
> >>> #include <rte_vect.h>
> >>> #if defined(__AVX__)
> >>> #define SCHED_VECTOR_ENABLE
> >>> #endif
> >>>
> >>> inside rte_sched.c
> >>>
> >>> Konstantin
> >>>
> >>>
> >>>>
> >>>>
> >>>>> 在 2015年11月27日，下午5:01，Thomas Monjalon <thomas.monjalon at 6wind.com> 写道：
> >>>>>
> >>>>> 2015-11-27 02:26, Qiu, Michael:
> >>>>>>>> On 2015/11/27 5:29, Thomas Monjalon wrote:
> >>>>>>>> 2015-11-26 18:49, Michael Qiu:
> >>>>>>>> gcc 4.3.4 does not include "immintrin.h", and will post below error:
> >>>>>>>>  lib/librte_sched/rte_sched.c:56:23: error:
> >>>>>>>>  immintrin.h: No such file or directory
> >>>>>>>>
> >>>>>>>> To avoid this issue, a gcc version check is need and a flag to indicate
> >>>>>>>> vector ablility.
> >>>>>>> [...]
> >>>>>>>> +#if (defined(__ICC) || (__GNUC__ == 4 &&  __GNUC_MINOR__ < 4))
> >>>>>>>> +
> >>>>>>>> +#if defined(__AVX__)
> >>>>>>>> #include <immintrin.h>
> >>>>>>>> +#define SCHED_VECTOR_ENABLE
> >>>>>>>> +#endif
> >>>>>>>> +
> >>>>>>>> +#else
> >>>>>>>> +
> >>>>>>>> +#include <x86intrin.h>
> >>>>>>>> +#define SCHED_VECTOR_ENABLE
> >>>>>>>> +
> >>>>>>>> +#endif
> >>>>>>> This kind of complication is managed by EAL.
> >>>>>>> I think we should include rte_vect.h.
> >>>>>>
> >>>>>> As I know here it needs a flag to identify whether the platform support
> >>>>>> AVX, if not it will not use it, so I don't know if we could only simply
> >>>>>> include rte_vect.h?
> >>>>>
> >>>>> It's not exclusive.
> >>>>> You can include rte_vect.h and check AVX to define SCHED_VECTOR_ENABLE.
> >>>>>