[FFmpeg-devel] [PATCH 1/7] lavc/me_cmp: R-V V pix_abs

[Rémi Denis-Courmont]

Le 9 février 2024 00:39:38 GMT+02:00, flow gg <hlefthleft at gmail.com> a écrit :
>From my understanding, to use larger group multipliers, one needs to
>utilize vlse64 (8x8) vlse128 (16x16).
>
>However, due to the use in tests of
>
>ptr = img2 + y * WIDTH + x;
>d2 = call_ref(NULL, img1, ptr, WIDTH, h);
>d1 = call_new(NULL, img1, ptr, WIDTH, h);
>
>will get:  pix_abs_1_0_rvv_i32 (fatal signal 7: Bus error)
>
>Because it can only load according to e8, it seems there's no way to use
>larger group multipliers.

vlse128.v requires 128-bit elements, which no hardware supports. vlse64.v works just fine; we're already using it. There's also the possibility of segmented strided loads, or simply multiple unit loads.

In any case, unrolling one way or other should improve performance.


>
>
>
>Rémi Denis-Courmont <remi at remlab.net> 于2024年2月9日周五 03:41写道：
>
>> Le keskiviikkona 7. helmikuuta 2024, 2.01.23 EET flow gg a écrit :
>> > I think in most cases it is like this, but specifically for this
>> function,
>> > using Reduction only once would be slower.
>> >
>> > The currently submitted version roughly takes:
>> > pix_abs_0_0_rvv_i32: 136.2
>> >
>> > The version that uses Reduction only once takes:
>> > pix_abs_0_0_rvv_i32: 169.2
>>
>> You're only using one vector and half a vector respectively, so the
>> logarithmic time of the sum is relatively small.
>>
>> But are you sure that it wouldn't be faster to process multiple rows and
>> larger group multiplers?
>>
>> > Here is the implementation of the version that uses it only once:
>> >
>> > func ff_pix_abs16_temp_rvv, zve32x
>> >         vsetivli        zero, 16, e32, m4, ta, ma
>> >         vmv.v.i         v24, 0
>> >         vmv.s.x         v0, zero
>> > 1:
>> >         vsetvli         zero, zero, e8, m1, tu, ma
>> >         vle8.v          v4, (a1)
>> >         vle8.v          v12, (a2)
>> >         addi            a4, a4, -1
>> >         vwsubu.vv       v16, v4, v12
>> >         add             a1, a1, a3
>> >         vwsubu.vv       v20, v12, v4
>> >         vsetvli         zero, zero, e16, m2, tu, ma
>> >         vmax.vv         v16, v16, v20
>> >         add             a2, a2, a3
>> >         vwadd.wv        v24, v24, v16
>> >         bnez            a4, 1b
>> >
>> >         vsetvli         zero, zero, e32, m4, ta, ma
>> >         vwredsumu.vs    v0, v24, v0
>> >         vmv.x.s         a0, v0
>> >         ret
>> > endfunc
>> >
>> > Rémi Denis-Courmont <remi at remlab.net> 于2024年2月7日周三 00:58写道：
>> >
>> > > Hi,
>> > >
>> > > To sum a vector, you should only reduce once at the end of the
>> function,
>> > > c.f.
>> > > how it's done in existing scalar products. Reduction instructions are
>> > > (intrinsically) slow.
>> > >
>> > > --
>> > > Rémi Denis-Courmont
>> > > http://www.remlab.net/
>> >
>> > _______________________________________________
>> > 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".
>>
>>
>> --
>> 雷米‧德尼-库尔蒙
>> http://www.remlab.net/
>>
>>
>>
>> _______________________________________________
>> 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".