[FFmpeg-devel] [PATCH v2] cpu: add a function for querying maximum required data alignment

[Hendrik Leppkes]

On Mon, Sep 4, 2017 at 10:53 PM, Michael Niedermayer
<michael at niedermayer.cc> wrote:
> On Mon, Sep 04, 2017 at 09:27:32PM +0200, wm4 wrote:
>> On Mon, 4 Sep 2017 21:18:35 +0200
>> Michael Niedermayer <michael at niedermayer.cc> wrote:
>>
>> > On Sat, Sep 02, 2017 at 09:47:38PM -0300, James Almer wrote:
>> > > From: Anton Khirnov <anton at khirnov.net>
>> > >
>> > > (cherry picked from commit e6bff23f1e11aefb16a2b5d6ee72bf7469c5a66e)
>> > > Signed-off-by: James Almer <jamrial at gmail.com>
>> > > ---
>> > > This is (afaics) the last API introduced to libav before the major bump.
>> > >
>> > > Now checking all the x86 flags that would require aligment of 16 bytes
>> > > or more.
>> > >
>> > >  doc/APIchanges      |  3 +++
>> > >  libavutil/cpu.c     | 35 +++++++++++++++++++++++++++++++++++
>> > >  libavutil/cpu.h     | 13 +++++++++++++
>> > >  libavutil/version.h |  2 +-
>> > >  4 files changed, 52 insertions(+), 1 deletion(-)
>> > >
>> > > diff --git a/doc/APIchanges b/doc/APIchanges
>> > > index 4effbf9364..6a57c210a9 100644
>> > > --- a/doc/APIchanges
>> > > +++ b/doc/APIchanges
>> > > @@ -15,6 +15,9 @@ libavutil:     2015-08-28
>> > >
>> > >  API changes, most recent first:
>> > >
>> > > +2017-09-xx - xxxxxxx - lavu 55.75.100 / lavu 55.31.0 - cpu.h
>> > > +  Add av_cpu_max_align() for querying maximum required data alignment.
>> > > +
>> > >  2017-09-01 - xxxxxxx - lavf 57.81.100 - avio.h
>> > >    Add avio_read_partial().
>> > >
>> > > diff --git a/libavutil/cpu.c b/libavutil/cpu.c
>> > > index a22da0fa8c..4f04da2460 100644
>> > > --- a/libavutil/cpu.c
>> > > +++ b/libavutil/cpu.c
>> > > @@ -16,9 +16,11 @@
>> > >   * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
>> > >   */
>> > >
>> > > +#include <stddef.h>
>> > >  #include <stdint.h>
>> > >  #include <stdatomic.h>
>> > >
>> > > +#include "attributes.h"
>> > >  #include "cpu.h"
>> > >  #include "cpu_internal.h"
>> > >  #include "config.h"
>> > > @@ -299,3 +301,36 @@ int av_cpu_count(void)
>> > >
>> > >      return nb_cpus;
>> > >  }
>> > > +
>> > > +size_t av_cpu_max_align(void)
>> > > +{
>> > > +    int av_unused flags = av_get_cpu_flags();
>> > > +
>> > > +#if ARCH_ARM || ARCH_AARCH64
>> > > +    if (flags & AV_CPU_FLAG_NEON)
>> > > +        return 16;
>> > > +#elif ARCH_PPC
>> > > +    if (flags & AV_CPU_FLAG_ALTIVEC)
>> > > +        return 16;
>> > > +#elif ARCH_X86
>> > > +    if (flags & (AV_CPU_FLAG_AVX2     |
>> > > +                 AV_CPU_FLAG_AVX      |
>> > > +                 AV_CPU_FLAG_FMA4     |
>> > > +                 AV_CPU_FLAG_FMA3))
>> > > +        return 32;
>> > > +    if (flags & (AV_CPU_FLAG_XOP      |
>> > > +                 AV_CPU_FLAG_AESNI    |
>> > > +                 AV_CPU_FLAG_SSE42    |
>> > > +                 AV_CPU_FLAG_SSE4     |
>> > > +                 AV_CPU_FLAG_SSSE3    |
>> > > +                 AV_CPU_FLAG_SSE3     |
>> > > +                 AV_CPU_FLAG_SSE2     |
>> > > +                 AV_CPU_FLAG_SSE      |
>> > > +                 AV_CPU_FLAG_AVXSLOW  |
>> > > +                 AV_CPU_FLAG_SSE3SLOW |
>> > > +                 AV_CPU_FLAG_SSE2SLOW))
>> > > +        return 16;
>> > > +#endif
>> > > +
>> > > +    return 8;
>> > > +}
>> > > diff --git a/libavutil/cpu.h b/libavutil/cpu.h
>> > > index de05593446..9e5d40affe 100644
>> > > --- a/libavutil/cpu.h
>> > > +++ b/libavutil/cpu.h
>> > > @@ -21,6 +21,8 @@
>> > >  #ifndef AVUTIL_CPU_H
>> > >  #define AVUTIL_CPU_H
>> > >
>> > > +#include <stddef.h>
>> > > +
>> > >  #include "attributes.h"
>> > >
>> > >  #define AV_CPU_FLAG_FORCE    0x80000000 /* force usage of selected flags (OR) */
>> >
>> >
>> > > @@ -113,4 +115,15 @@ int av_parse_cpu_caps(unsigned *flags, const char *s);
>> > >   */
>> > >  int av_cpu_count(void);
>> > >
>> > > +/**
>> > > + * Get the maximum data alignment that may be required by FFmpeg.
>> > > + *
>> > > + * Note that this is affected by the build configuration and the CPU flags mask,
>> > > + * so e.g. if the CPU supports AVX, but libavutil has been built with
>> > > + * --disable-avx or the AV_CPU_FLAG_AVX flag has been disabled through
>> > > + *  av_set_cpu_flags_mask(), then this function will behave as if AVX is not
>> > > + *  present.
>> > > + */
>> > > +size_t av_cpu_max_align(void);
>> >
>> > This might interact badly with runtime cpu flags/mask changes
>> >
>> > If its used to choose the alignment for allocated frames and
>> > after some are allocated the cpu flags are changed there could be
>> > still frames in queues that may not have sufficient alignment for the
>> > new flags
>>
>> There's no such thing as runtime CPU flag changes.
>
> We even have an API to change the cpu flags at runtime.
>
> av_set_cpu_flags_mask() and av_force_cpu_flags()
>
> There is no restriction in the API on when they can be called.
>
> you can call av_force_cpu_flags(0) then open a decoder then call
> av_force_cpu_flags(AV_CPU_FLAG_MMX|AV_CPU_FLAG_SSE)
> then open a filter or encoder
> then run the code and it could crash as the allocated frames from
> earlier are not aligned enough for the 2nd filter or encoder
>
> There also may be other scenarios where this can occur
>

Thats why API to change flags is a dumb feature to have user-facing.
It should correspond to hardware support and nothing else.

- Hendrik