[FFmpeg-devel] libswscale.c : ff_xyz12Torgb48 expensive unaligned 16 byte accesses

[FFmpeg-devel] libswscale.c : ff_xyz12Torgb48 expensive unaligned 16 byte accesses

[Chitra Dey Sarkar via ffmpeg-devel]

H
We have been profiling FFmpeg at Microsoft and have identified that ff_xyz12ToRgb48 has a high sample count ( profiled every 1ms )

It seems like ff_xyz12ToRgb48 has performance penalty for

  1.  Unaligned read and write access
  2.  Access to xyz2rgb_matrix
  3.  Multiplication

I would be interested in optimizing this code , wanted to check if there is an existing optimized version of this function, or any recommended approach to improve it(?
I can move the repeated access to xyz2rgb_matrix outside the inner loop and load a full cache line at once to extract the X, Y, and Z values more efficiently-but I wanted to start by getting some initial feedback or thoughts before proceeding further


File : FFmpeg/libswscale/swscale.c at a79720e10f30e9fd18bd78242ce96dde06461343 * FFmpeg/FFmpeg<https://github.com/FFmpeg/FFmpeg/blob/a79720e10f30e9fd18bd78242ce96dde06461343/libswscale/swscale.c#L739>

void ff_xyz12Torgb48(const SwsInternal *c, uint8_t *dst, int dst_stride,
                     const uint8_t *src, int src_stride, int w, int h)
{
.......... Unaligned read .......................
               x = AV_RL16(src16 + xp + 0);
                y = AV_RL16(src16 + xp + 1);
                z = AV_RL16(src16 + xp + 2);


.......... DRAM Access and multiply  .......................

            // convert from XYZlinear to sRGBlinear
            r = c->xyz2rgb_matrix[0][0] * x +
                c->xyz2rgb_matrix[0][1] * y +
                c->xyz2rgb_matrix[0][2] * z >> 12;
            g = c->xyz2rgb_matrix[1][0] * x +
                c->xyz2rgb_matrix[1][1] * y +
                c->xyz2rgb_matrix[1][2] * z >> 12;
            b = c->xyz2rgb_matrix[2][0] * x +
                c->xyz2rgb_matrix[2][1] * y +
                c->xyz2rgb_matrix[2][2] * z >> 12;

.......... RMW Access .......................
                AV_WL16(dst16 + xp + 0, c->rgbgamma[r] << 4);
                AV_WL16(dst16 + xp + 1, c->rgbgamma[g] << 4);
                AV_WL16(dst16 + xp + 2, c->rgbgamma[b] << 4);


Regards
Chitra
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Re: [FFmpeg-devel] libswscale.c : ff_xyz12Torgb48 expensive unaligned 16 byte accesses

[James Almer]

[-- Attachment #1.1.1: Type: text/plain, Size: 1002 bytes --]

On 5/25/2025 9:44 PM, Chitra Dey Sarkar via ffmpeg-devel wrote:
> H
> We have been profiling FFmpeg at Microsoft and have identified that ff_xyz12ToRgb48 has a high sample count ( profiled every 1ms )
> 
> It seems like ff_xyz12ToRgb48 has performance penalty for
> 
>    1.  Unaligned read and write access
>    2.  Access to xyz2rgb_matrix
>    3.  Multiplication
> 
> I would be interested in optimizing this code , wanted to check if there is an existing optimized version of this function, or any recommended approach to improve it(?
> I can move the repeated access to xyz2rgb_matrix outside the inner loop and load a full cache line at once to extract the X, Y, and Z values more efficiently-but I wanted to start by getting some initial feedback or thoughts before proceeding further

Niklas is currently working in a library rewrite, which may affect xyz 
scaling. But in any case, the obvious approach is a SIMD optimized 
implementation (SSE2, AVX2, etc) of this function.


[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]

[-- Attachment #2: Type: text/plain, Size: 251 bytes --]

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".