[FFmpeg-devel] [PATCH] swscale/aarch64/rgb2rgb: Implemented NEON shuf routines

[Martin Storsjö]

On Tue, 28 Jan 2025, Krzysztof Pyrkosz via ffmpeg-devel wrote:

> The key idea is to pass the pre-generated tables to the TBL instruction
> and churn through the data 16 bytes at a time. The remaining 4 elements
> are handled with a specialized block located at the end of the routine.
>
> The 3210 variant can be implemented using rev32, but surprisingly it is
> slower than the generic TBL on A78, but much faster on A72. I wrapped it
> in #if 0 block.

So the tradeoff is essentially this:

A78:
shuffle_bytes_3210_c:                                  138.0 ( 1.00x)
shuffle_bytes_3210_tbl_neon:                            22.0 ( 6.27x)
shuffle_bytes_3210_rev32_neon:                          28.5 ( 4.88x)

A72:
shuffle_bytes_3210_c:                                  195.8 ( 1.00x)
shuffle_bytes_3210_tbl_neon:                            37.8 ( 5.19x)
shuffle_bytes_3210_rev32_neon:                          30.8 ( 6.33x)

Yeah it doesn't really make much of a difference which on we pick here; 
we're much faster than the C code in any case. I guess favouring tbl for 
the newer cores is the right choice to make.

> diff --git a/libswscale/aarch64/rgb2rgb_neon.S b/libswscale/aarch64/rgb2rgb_neon.S
> index 1382e00261..a69a211ad4 100644
> --- a/libswscale/aarch64/rgb2rgb_neon.S
> +++ b/libswscale/aarch64/rgb2rgb_neon.S
> @@ -296,3 +359,99 @@ function ff_deinterleave_bytes_neon, export=1
> 0:
>         ret
> endfunc
> +
> +.macro neon_shuf shuf
> +function ff_shuffle_bytes_\shuf\()_neon, export=1
> +        movrel          x9, shuf_\shuf\()_tbl
> +        ld1             {v1.16b}, [x9]
> +        and             w5, w2, #~15
> +        and             w3, w2, #8
> +        and             w4, w2, #4
> +        cbz             w5, 2f
> +1:
> +        subs            w5, w5, #16
> +        ld1             {v0.16b}, [x0], #16
> +        tbl             v0.16b, {v0.16b}, v1.16b
> +        st1             {v0.16b}, [x1], #16
> +        b.gt            1b

By moving the subs to after the ld1, on the Cortex A53, I get the runtime 
lowered from this:

shuffle_bytes_0321_c:                                  283.0 ( 1.00x)
shuffle_bytes_0321_neon:                                68.0 ( 4.16x)

to this:

shuffle_bytes_0321_neon:                                60.8 ( 4.66x)

So I'm squashing such a change into it.


> +#
> +#if 0
> +function ff_shuffle_bytes_3210_neon, export=1
> +        and             w5, w2, #~(15)

While it is nice to keep this as reference, it's kinda dead code here, so 
I would suggest we just drop it for now. Good that you investigated it 
though!

Other than that, this looks really good, so I'll push it with those 
changes.

// Martin