Re: [FFmpeg-devel] [PATCH] swscale/aarch64: dotprod implementation of rgba32_to_Y

[Niklas Haas <ffmpeg@haasn.xyz>]

On Fri, 28 Feb 2025 10:31:19 +0800 Zhao Zhili <quinkblack@foxmail.com> wrote:
> Cc haasn.
>
> Libswscale in under refactor. Does current asm works after refactor, or they need to be refactored or
> rewrite after? If it’s the second case, maybe we should hold on to do more asm with libswscale
> before hassn work done.

No, almost all current asm will be unused after the rewrite. There are some we
can in theory reuse, but for the most part, it doesn't seem to be worth it.

Especially for the very bespoke functions like this one.

For context, in general, the focus in nu-swscale is to focus more on smaller,
flexible primitives and have the calling code combine them as needed. So instead
of a "brga_to_y" function, you would have a sequence that looks like this:

Operation list:
  [ u8 XXXX -> dddX] SWS_OP_READ         : 4 elem(s) packed >> 0
  [ u8 ...X -> dddX] SWS_OP_SWIZZLE      : 2103
  [ u8 ...X -> dddX] SWS_OP_CONVERT      : u8 -> f32
  [f32 ...X -> .XXX] SWS_OP_LINEAR       : dot3 [[0.299000 0.587000 0.114000 0 0] [0 1 0 0 0] [0 0 1 0 0] [0 0 0 1 0]]
  [f32 .XXX -> .XXX] SWS_OP_DITHER       : 16x16 {255 _ _ _}
  [f32 .XXX -> dXXX] SWS_OP_CONVERT      : f32 -> u8
  [ u8 .XXX -> XXXX] SWS_OP_WRITE        : 1 elem(s) packed >> 0

Where each low-level implementation can combine one, or multiple, such
operations together. For example, in the current prototype, SWS_OP_CONVERT and
SWS_OP_WRITE can be fused together into a single implementation.

Note also the conversion to float. I found that the cost of going through
floats seems to be lower on average, across all tested platforms, than the
extra cost of dealing with integers (which require extra shifting, extra
dithering, and extra width conversions - all of which exceed the cost of just
one extra float->int conversion step). This also comes with improved accuracy.

>
> > On Feb 28, 2025, at 06:44, Krzysztof Pyrkosz via ffmpeg-devel <ffmpeg-devel@ffmpeg.org> wrote:
> >
> > ---
> > I was curious whether it's possible to implement this function without
> > any widening, and it turns out it not only is, but it's quite
> > performant at the same time!
> >
> > The idea is to split the 16 bit coefficients into lower and upper half,
> > invoke udot for the lower half, shift by 8, and follow by udot for the
> > upper half. The code is based upon existing version.
> >
> > Benchmark on A78:
> > bgra_to_y_128_c:                                       682.0 ( 1.00x)
> > bgra_to_y_128_neon:                                    181.2 ( 3.76x)
> > bgra_to_y_128_dotprod:                                 117.8 ( 5.79x)
> > bgra_to_y_1080_c:                                     5742.5 ( 1.00x)
> > bgra_to_y_1080_neon:                                  1472.5 ( 3.90x)
> > bgra_to_y_1080_dotprod:                                906.5 ( 6.33x)
> > bgra_to_y_1920_c:                                    10194.0 ( 1.00x)
> > bgra_to_y_1920_neon:                                  2589.8 ( 3.94x)
> > bgra_to_y_1920_dotprod:                               1573.8 ( 6.48x)
> >
> > Krzysztof
> >
> > libswscale/aarch64/input.S   | 88 ++++++++++++++++++++++++++++++++++++
> > libswscale/aarch64/swscale.c | 17 +++++++
> > 2 files changed, 105 insertions(+)
> >
> > diff --git a/libswscale/aarch64/input.S b/libswscale/aarch64/input.S
> > index 5cb18711fb..5fe6c3f6f5 100644
> > --- a/libswscale/aarch64/input.S
> > +++ b/libswscale/aarch64/input.S
> > @@ -313,3 +313,91 @@ rgbToUV_neon bgr24, rgb24, element=3
> > rgbToUV_neon bgra32, rgba32, element=4
> >
> > rgbToUV_neon abgr32, argb32, element=4, alpha_first=1
> > +
> > +#if HAVE_DOTPROD
> > +ENABLE_DOTPROD
> > +
> > +function ff_bgra32ToY_neon_dotprod, export=1
> > +        cmp             w4, #0                  // check width > 0
> > +        ldp             w12, w11, [x5]          // w12: ry, w11: gy
> > +        ldr             w10, [x5, #8]           // w10: by
> > +        b.gt            4f
> > +        ret
> > +endfunc
> > +
> > +function ff_rgba32ToY_neon_dotprod, export=1
> > +        cmp             w4, #0                  // check width > 0
> > +        ldp             w10, w11, [x5]          // w10: ry, w11: gy
> > +        ldr             w12, [x5, #8]           // w12: by
> > +        b.le            3f
> > +4:
> > +        mov             w9, #256                // w9 = 1 << (RGB2YUV_SHIFT - 7)
> > +        movk            w9, #8, lsl #16         // w9 += 32 << (RGB2YUV_SHIFT - 1)
> > +        dup             v6.4s, w9               // w9: const_offset
> > +
> > +        cmp             w4, #16
> > +        mov             w7, w10
> > +        bfi             w7, w11, 8, 8
> > +        bfi             w7, w12, 16, 8
> > +        dup             v0.4s, w7
> > +
> > +        lsr             w6, w10, #8
> > +        lsr             w7, w11, #8
> > +        lsr             w8, w12, #8
> > +
> > +        bfi             w6, w7, 8, 8
> > +        bfi             w6, w8, 16, 8
> > +        dup             v1.4s, w6
> > +        b.lt            2f
> > +1:
> > +        ld1             { v16.16b, v17.16b, v18.16b, v19.16b }, [x1], #64
> > +        sub             w4, w4, #16             // width -= 16
> > +        cmp             w4, #16                 // width >= 16 ?
> > +
> > +        mov             v2.16b, v6.16b
> > +        mov             v3.16b, v6.16b
> > +        mov             v4.16b, v6.16b
> > +        mov             v5.16b, v6.16b
> > +
> > +        udot            v2.4s, v16.16b, v0.16b
> > +        udot            v3.4s, v17.16b, v0.16b
> > +        udot            v4.4s, v18.16b, v0.16b
> > +        udot            v5.4s, v19.16b, v0.16b
> > +
> > +        ushr            v2.4s, v2.4s, #8
> > +        ushr            v3.4s, v3.4s, #8
> > +        ushr            v4.4s, v4.4s, #8
> > +        ushr            v5.4s, v5.4s, #8
> > +
> > +        udot            v2.4s, v16.16b, v1.16b
> > +        udot            v3.4s, v17.16b, v1.16b
> > +        udot            v4.4s, v18.16b, v1.16b
> > +        udot            v5.4s, v19.16b, v1.16b
> > +
> > +        sqshrn          v16.4h, v2.4s, #1
> > +        sqshrn2         v16.8h, v3.4s, #1
> > +        sqshrn          v17.4h, v4.4s, #1
> > +        sqshrn2         v17.8h, v5.4s, #1
> > +
> > +        stp             q16, q17, [x0], #32     // store to dst
> > +        b.ge            1b
> > +        cbz             x4, 3f
> > +2:
> > +        ldrb            w13, [x1]               // w13: r
> > +        ldrb            w14, [x1, #1]           // w14: g
> > +        ldrb            w15, [x1, #2]           // w15: b
> > +
> > +        smaddl          x13, w13, w10, x9       // x13 = ry * r + const_offset
> > +        smaddl          x13, w14, w11, x13      // x13 += gy * g
> > +        smaddl          x13, w15, w12, x13      // x13 += by * b
> > +        asr             w13, w13, #9            // x13 >>= 9
> > +        sub             w4, w4, #1              // width--
> > +        add             x1, x1, #4
> > +        strh            w13, [x0], #2           // store to dst
> > +        cbnz            w4, 2b
> > +3:
> > +        ret
> > +endfunc
> > +
> > +DISABLE_DOTPROD
> > +#endif
> > diff --git a/libswscale/aarch64/swscale.c b/libswscale/aarch64/swscale.c
> > index 92c49dcf3a..ffcc6a0605 100644
> > --- a/libswscale/aarch64/swscale.c
> > +++ b/libswscale/aarch64/swscale.c
> > @@ -210,6 +210,9 @@ void ff_##name##ToUV_neon(uint8_t *, uint8_t *, const uint8_t *, \
> > void ff_##name##ToUV_half_neon(uint8_t *, uint8_t *, const uint8_t *, \
> >                               const uint8_t *, const uint8_t *, int w, \
> >                               uint32_t *coeffs, void *)
> > +#define NEON_INPUT_DOTPROD(name) \
> > +void ff_##name##ToY_neon_dotprod(uint8_t *dst, const uint8_t *src, const uint8_t *, \
> > +                                 const uint8_t *, int w, uint32_t *coeffs, void *);
> >
> > NEON_INPUT(abgr32);
> > NEON_INPUT(argb32);
> > @@ -217,6 +220,8 @@ NEON_INPUT(bgr24);
> > NEON_INPUT(bgra32);
> > NEON_INPUT(rgb24);
> > NEON_INPUT(rgba32);
> > +NEON_INPUT_DOTPROD(bgra32);
> > +NEON_INPUT_DOTPROD(rgba32);
> >
> > void ff_lumRangeFromJpeg8_neon(int16_t *dst, int width,
> >                                uint32_t coeff, int64_t offset);
> > @@ -295,6 +300,12 @@ av_cold void ff_sws_init_swscale_aarch64(SwsInternal *c)
> >                 c->chrToYV12 = ff_bgr24ToUV_neon;
> >             break;
> >         case AV_PIX_FMT_BGRA:
> > +#if HAVE_DOTPROD
> > +            if (have_dotprod(cpu_flags)) {
> > +                c->lumToYV12 = ff_bgra32ToY_neon_dotprod;
> > +            }
> > +	    else
> > +#endif
> >             c->lumToYV12 = ff_bgra32ToY_neon;
> >             if (c->chrSrcHSubSample)
> >                 c->chrToYV12 = ff_bgra32ToUV_half_neon;
> > @@ -309,6 +320,12 @@ av_cold void ff_sws_init_swscale_aarch64(SwsInternal *c)
> >                 c->chrToYV12 = ff_rgb24ToUV_neon;
> >             break;
> >         case AV_PIX_FMT_RGBA:
> > +#if HAVE_DOTPROD
> > +            if (have_dotprod(cpu_flags)) {
> > +                c->lumToYV12 = ff_rgba32ToY_neon_dotprod;
> > +            }
> > +	    else
> > +#endif
> >             c->lumToYV12 = ff_rgba32ToY_neon;
> >             if (c->chrSrcHSubSample)
> >                 c->chrToYV12 = ff_rgba32ToUV_half_neon;
> > --
> > 2.47.2
> >
> > _______________________________________________
> > 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".
>
_______________________________________________
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".