[FFmpeg-devel] [PATCH] lavc/h264chroma: RISC-V V add motion compensation for 8x8 chroma blocks

[Arnie Chang]

On Sat, May 20, 2023 at 1:12 AM Rémi Denis-Courmont <remi at remlab.net> wrote:

> > +    li    t4, 0
> > +    li    t2, 0
> > +    addi    a5, t3, 1
> > +    slli    t3, a2, 2
> > +.LBB0_3:                                # if (xy != 0)
> > +    add    a4, a1, t4
> > +    vsetvli    zero, a5, e8, m1, ta, ma
> > +    addiw    t2, t2, 4
> > +    vle8.v    v10, (a4)
> > +    add    a4, a4, a2
> > +    vslidedown.vi    v11, v10, 1
>
> Isn't vslide1down.vx zero potentially faster than vslidedown.vi 1?
>

It depends on hardware design,
but in general, vslide1down.vi is typically not slower than vslidedown.vx
Using vslide1down.vi would be better here, I will fix it.



> > +    vsetivli    zero, 8, e8, m1, ta, ma
>
> Do we really need to reconfigure the active vector length so many times? I
> suspect that is not going to go down to well with some implementations.
>

We need to reconfigure it because the VL is changed.
The VL for vslidedown differs from that of the other instructions.

> +    vwmaccu.vx    v12, t1, v15
> > +    vwmaccu.vx    v16, a7, v15
> > +    vsetvli    zero, a5, e8, m1, ta, ma
> > +    vle8.v    v14, (a4)
> > +    vsetivli    zero, 8, e8, m1, ta, ma
> > +    add    a4, a0, t4
> > +    add    t4, t4, t3
>
> I could be totally wrong since I have no hardware to verify with, but I
> would
> assume that it is preferable to interleave independent scalar and vector
> instructions whence possible. For out-of-order processors, it shouldn't
> matter, but I suppose that it would on in-order multi-issue processors.
>

Interleaving those instructions can improve overall performance