[FFmpeg-devel] [PATCH 5/5] lavc/aarch64: Provide neon implementation of nsse16

[Martin Storsjö]

On Tue, 6 Sep 2022, Hubert Mazur wrote:

> Add vectorized implementation of nsse16 function.
>
> Performance comparison tests are shown below.
> - nsse_0_c: 707.0
> - nsse_0_neon: 120.0
>
> Benchmarks and tests run with checkasm tool on AWS Graviton 3.
>
> Signed-off-by: Hubert Mazur <hum at semihalf.com>
> ---
> libavcodec/aarch64/me_cmp_init_aarch64.c |  15 +++
> libavcodec/aarch64/me_cmp_neon.S         | 124 +++++++++++++++++++++++
> 2 files changed, 139 insertions(+)
>
> diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
> index 8c295d5457..ea7f295373 100644
> --- a/libavcodec/aarch64/me_cmp_init_aarch64.c
> +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
> @@ -49,6 +49,10 @@ int vsse16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
>                 ptrdiff_t stride, int h);
> int vsse_intra16_neon(MpegEncContext *c, const uint8_t *s, const uint8_t *dummy,
>                       ptrdiff_t stride, int h);
> +int nsse16_neon(int multiplier, const uint8_t *s, const uint8_t *s2,
> +                ptrdiff_t stride, int h);
> +int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
> +                        ptrdiff_t stride, int h);
> 
> av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
> {
> @@ -72,5 +76,16 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
>
>         c->vsse[0] = vsse16_neon;
>         c->vsse[4] = vsse_intra16_neon;
> +
> +        c->nsse[0] = nsse16_neon_wrapper;
>     }
> }
> +
> +int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
> +                        ptrdiff_t stride, int h)
> +{
> +        if (c)
> +            return nsse16_neon(c->avctx->nsse_weight, s1, s2, stride, h);
> +        else
> +            return nsse16_neon(8, s1, s2, stride, h);
> +}

The body of this function is still indented 4 spaces too much.

> diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
> index cf2b8da425..bd21122a21 100644
> --- a/libavcodec/aarch64/me_cmp_neon.S
> +++ b/libavcodec/aarch64/me_cmp_neon.S
> @@ -847,3 +847,127 @@ function vsse_intra16_neon, export=1
>
>         ret
> endfunc
> +
> +function nsse16_neon, export=1
> +        // x0           multiplier
> +        // x1           uint8_t *pix1
> +        // x2           uint8_t *pix2
> +        // x3           ptrdiff_t stride
> +        // w4           int h
> +
> +        str             x0, [sp, #-0x40]!
> +        stp             x1, x2, [sp, #0x10]
> +        stp             x3, x4, [sp, #0x20]
> +        str             x30, [sp, #0x30]
> +        bl              X(sse16_neon)
> +        ldr             x30, [sp, #0x30]
> +        mov             w9, w0                                  // here we store score1
> +        ldr             x5, [sp]
> +        ldp             x1, x2, [sp, #0x10]
> +        ldp             x3, x4, [sp, #0x20]
> +        add             sp, sp, #0x40
> +
> +        movi            v16.8h, #0
> +        movi            v17.8h, #0
> +        movi            v18.8h, #0
> +        movi            v19.8h, #0
> +
> +        ld1             {v0.16b}, [x1], x3
> +        subs            w4, w4, #1                              // we need to make h-1 iterations
> +        ext             v1.16b, v0.16b, v0.16b, #1              // x1 + 1
> +        ld1             {v2.16b}, [x2], x3
> +        cmp             w4, #2
> +        ext             v3.16b, v2.16b, v2.16b, #1              // x2 + 1

The scheduling here (and in both loops below) is a bit non-ideal; we use 
v0 very early after loading it, while we could start the load of v2 before 
that.

> +
> +        b.lt            2f
> +
> +// make 2 iterations at once
> +1:
> +        ld1             {v4.16b}, [x1], x3
> +        ld1             {v20.16b}, [x1], x3

The load of v20 here would stall a bit, since the first load updates x1. 
Instead we can launch the load of v6 inbetween these two

> +        ext             v5.16b, v4.16b, v4.16b, #1              // x1 + stride + 1
> +        ext             v21.16b, v20.16b, v20.16b, #1

We can move the use of v20 a bit later here, inbetween the two other 
exts.

These changes (plus a similar minor one to the non-unrolled version at the 
bottom) produces the following benchmark change for me:

Before:   Cortex A53     A72     A73
nsse_0_neon:   401.0   198.0   194.5
After:
nsse_0_neon:   377.0   198.7   196.5

(The differences on A72 and A73 are within the measurement noise, those 
numbers vary more than that from one run to another.)

You can squash in the attached patch for simplicity. Also, remember to fix 
the indentation of the wrapper function.

With those changes, plus updated benchmark numbers in the commit messages, 
I think this patchset should be good to go.

// Martin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-squash-nsse16-Tune-scheduling.patch
Type: text/x-diff
Size: 2496 bytes
Desc: 
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20220907/ccfbd3d8/attachment.patch>