[FFmpeg-devel] [PATCH 1/2] swscale: rgb_to_yuv neon optimizations

[Martin Storsjö]

On Sat, 31 May 2025, Dmitriy Kovalenko wrote:

> I've found quite a few ways to optimize existing ffmpeg's rgb to yuv
> subsampled conversion. In this patch stack I'll try to
> improve the perofrmance.
>
> This particular set of changes is a small improvement to all the
> existing functions and macro. The biggest performance gain is
> coming from post loading increment of the pointer and immediate
> ~~prefetching of the memory blocks~~(was moved to the next patch in the stack) and interleaving the multiplication shifting operations of
> different registers for better scheduling.

Why keep the mention of prefetching here, when it no longer is included in 
the patch at all? This is what you suggest is encoded as the final, 
immutable commit message describing this change.

I have further inline comments below, please read them all.

> Also changed a bunch of places where cmp + b.le was used instead
> of one instruction cbnz/tbnz and some other small cleanups.
>
> Here are checkasm results on the macbook pro with the latest M4 max
>
> <before>
>
> bgra_to_uv_1080_c:                                     257.5 ( 1.00x)
> bgra_to_uv_1080_neon:                                  211.9 ( 1.22x)
> bgra_to_uv_1920_c:                                     467.1 ( 1.00x)
> bgra_to_uv_1920_neon:                                  379.3 ( 1.23x)
> bgra_to_uv_half_1080_c:                                198.9 ( 1.00x)
> bgra_to_uv_half_1080_neon:                             125.7 ( 1.58x)
> bgra_to_uv_half_1920_c:                                346.3 ( 1.00x)
> bgra_to_uv_half_1920_neon:                             223.7 ( 1.55x)
>
> <after>
>
> bgra_to_uv_1080_c:                                     268.3 ( 1.00x)
> bgra_to_uv_1080_neon:                                  176.0 ( 1.53x)
> bgra_to_uv_1920_c:                                     456.6 ( 1.00x)
> bgra_to_uv_1920_neon:                                  307.7 ( 1.48x)
> bgra_to_uv_half_1080_c:                                193.2 ( 1.00x)
> bgra_to_uv_half_1080_neon:                              96.8 ( 2.00x)
> bgra_to_uv_half_1920_c:                                347.2 ( 1.00x)
> bgra_to_uv_half_1920_neon:                             182.6 ( 1.92x)
>
> With my proprietary test on IOS it gives around 70% of performance
> improvement converting bgra 1920x1920 image to yuv420p
>
> On my linux arm cortex-r processing the performance improvement not that
> visible but still consistently faster by 5-10% than the current
> implementation.
> ---
> libswscale/aarch64/input.S | 143 +++++++++++++++++++++++--------------
> 1 file changed, 91 insertions(+), 52 deletions(-)

> @@ -292,7 +330,7 @@ function ff_\fmt_rgb\()ToUV_neon, export=1
>         smaddl          x8, w16, w10, x9        // x8 = ru * r + const_offset
>         smaddl          x8, w17, w11, x8        // x8 += gu * g
>         smaddl          x8, w4, w12, x8         // x8 += bu * b
> -        asr             w8, w8, #9              // x8 >>= 9
> +        asr             x8, x8, #9              // x8 >>= 9
>         strh            w8, [x0], #2            // store to dst_u
>

Here you _still_ have one instance of these unrelated changes left in your 
patch.

>         smaddl          x8, w16, w13, x9        // x8 = rv * r + const_offset
> @@ -401,3 +439,4 @@ endfunc
>
> DISABLE_DOTPROD
> #endif
> +
> --

Here you are adding one unrelated empty line at the end of the file. Don't 
include any unrelated changes in your patches.

Before sending a patch, do review it yourself first, checking for any such 
unrelated stray changes.

Other than those details, the rest of the patch looks ok.

// Martin