[FFmpeg-devel] [PATCH 1/2] libavutil/cpu: Add AV_CPU_FLAG_SLOW_GATHER.

[Alan Kelly]

On Mon, Dec 20, 2021 at 3:53 PM James Almer <jamrial at gmail.com> wrote:

>
>
> On 12/20/2021 11:47 AM, Lynne wrote:
> > 20 Dec 2021, 15:43 by alankelly-at-google.com at ffmpeg.org:
> >
> >> This flag is set on Haswell and earlier and all AMD cpus.
> >> ---
> >>   Removes unnecessary indentation, clarifies comment and only sets flag
> on AMD
> >>   cpus with AVX2.
> >>   libavutil/cpu.h     |  1 +
> >>   libavutil/x86/cpu.c | 14 +++++++++++++-
> >>   2 files changed, 14 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/libavutil/cpu.h b/libavutil/cpu.h
> >> index ae443eccad..ce9bf14bf7 100644
> >> --- a/libavutil/cpu.h
> >> +++ b/libavutil/cpu.h
> >> @@ -54,6 +54,7 @@
> >>   #define AV_CPU_FLAG_BMI1        0x20000 ///< Bit Manipulation
> Instruction Set 1
> >>   #define AV_CPU_FLAG_BMI2        0x40000 ///< Bit Manipulation
> Instruction Set 2
> >>   #define AV_CPU_FLAG_AVX512     0x100000 ///< AVX-512 functions:
> requires OS support even if YMM/ZMM registers aren't used
> >> +#define AV_CPU_FLAG_SLOW_GATHER  0x2000000 ///< CPU has slow gathers.
> >>
> >>   #define AV_CPU_FLAG_ALTIVEC      0x0001 ///< standard
> >>   #define AV_CPU_FLAG_VSX          0x0002 ///< ISA 2.06
> >> diff --git a/libavutil/x86/cpu.c b/libavutil/x86/cpu.c
> >> index bcd41a50a2..563984f234 100644
> >> --- a/libavutil/x86/cpu.c
> >> +++ b/libavutil/x86/cpu.c
> >> @@ -146,8 +146,16 @@ int ff_get_cpu_flags_x86(void)
> >>   if (max_std_level >= 7) {
> >>   cpuid(7, eax, ebx, ecx, edx);
> >>   #if HAVE_AVX2
> >> -        if ((rval & AV_CPU_FLAG_AVX) && (ebx & 0x00000020))
> >> +        if ((rval & AV_CPU_FLAG_AVX) && (ebx & 0x00000020)) {
> >>   rval |= AV_CPU_FLAG_AVX2;
> >> +            cpuid(1, eax, ebx, ecx, std_caps);
> >> +            family = ((eax >> 8) & 0xf) + ((eax >> 20) & 0xff);
> >> +            model  = ((eax >> 4) & 0xf) + ((eax >> 12) & 0xf0);
> >> +            /* Haswell has slow gather */
> >> +            if(family == 6 && model < 70)
> >> +                rval |= AV_CPU_FLAG_SLOW_GATHER;
> >> +        }
> >> +
> >>   #if HAVE_AVX512 /* F, CD, BW, DQ, VL */
> >>   if ((xcr0_lo & 0xe0) == 0xe0) { /* OPMASK/ZMM state */
> >>   if ((rval & AV_CPU_FLAG_AVX2) && (ebx & 0xd0030000) == 0xd0030000)
> >> @@ -196,6 +204,10 @@ int ff_get_cpu_flags_x86(void)
> >>   used unless explicitly disabled by checking AV_CPU_FLAG_AVXSLOW. */
> >>   if ((family == 0x15 || family == 0x16) && (rval & AV_CPU_FLAG_AVX))
> >>   rval |= AV_CPU_FLAG_AVXSLOW;
> >> +
> >> +        /* AMD cpus have slow gather */
> >> +            if(rval & AV_CPU_FLAG_AVX2)
> >> +                rval |= AV_CPU_FLAG_SLOW_GATHER;
> >>   }
> >>
> >
> > No, I'd rather limit AMD CPUs to all currently released CPUs.
> > Future ones are getting AVX512, which did speed up gathers on
> > Intel CPUs, as the ISA extension extended gathers and addded
> > scatters.
>
> I wouldn't hold my breath for that, but it's probably a good idea
> anyway. A check so it's flagged only on Excavator and Zen <= 3.
>
> >
> > Also your previous patch introduces ff_shuffle_filter_coefficients()
> > which is so bad it pretty much needs a complete rewrite.
> > You're also not detecting malloc errors or propagating them back.
>
> That's unrelated to this patch.
>
> >
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >
> > To unsubscribe, visit link above, or email
> > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>

Updated patch sent with check for family <= 25 so that future CPUs will
have avx2 hscale enabled by default.

I may have time this week to look at ff_shuffle_filter_coefficients.