On Mon, 26 Sep 2022, Grzegorz Bernacki wrote:

> Provide optimized implementation for vsse_intra8 for arm64.
>
> Performance tests are shown below.
> - vsse_5_c: 87.7
> - vsse_5_neon: 26.2
>
> Benchmarks and tests are run with checkasm tool on AWS Graviton 3.
> ---
> libavcodec/aarch64/me_cmp_init_aarch64.c |  4 ++
> libavcodec/aarch64/me_cmp_neon.S         | 53 ++++++++++++++++++++++++
> 2 files changed, 57 insertions(+)
>
> diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
> index f247372c94..defec37478 100644
> --- a/libavcodec/aarch64/me_cmp_init_aarch64.c
> +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
> @@ -74,6 +74,9 @@ int nsse8_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
> int vsse8_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
>                ptrdiff_t stride, int h);
> 
> +int vsse_intra8_neon(MpegEncContext *c, const uint8_t *s, const uint8_t *dummy,
> +                     ptrdiff_t stride, int h);
> +
> av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
> {
>     int cpu_flags = av_get_cpu_flags();
> @@ -102,6 +105,7 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
>         c->vsse[1] = vsse8_neon;
>
>         c->vsse[4] = vsse_intra16_neon;
> +        c->vsse[5] = vsse_intra8_neon;
>
>         c->nsse[0] = nsse16_neon_wrapper;
>         c->nsse[1] = nsse8_neon_wrapper;
> diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
> index 386d2de0c5..82ff05d3f0 100644
> --- a/libavcodec/aarch64/me_cmp_neon.S
> +++ b/libavcodec/aarch64/me_cmp_neon.S
> @@ -1111,6 +1111,59 @@ function vsse_intra16_neon, export=1
>         ret
> endfunc
> 
> +function vsse_intra8_neon, export=1
> +        // x0           unused
> +        // x1           uint8_t *pix1
> +        // x2           uint8_t *dummy
> +        // x3           ptrdiff_t stride
> +        // w4           int h
> +
> +        ld1             {v0.8b}, [x1], x3
> +        movi            v16.4s, #0
> +
> +        sub             w4, w4, #1 // we need to make h-1 iterations
> +        cmp             w4, #3
> +        b.lt            2f
> +
> +1:
> +        // v = abs( pix1[0] - pix1[0 + stride] )
> +        // score = sum( v * v )
> +        ld1             {v1.8b}, [x1], x3
> +        ld1             {v2.8b}, [x1], x3
> +        uabd            v30.8b, v0.8b, v1.8b
> +        ld1             {v3.8b}, [x1], x3
> +        umull           v29.8h, v30.8b, v30.8b
> +        uabd            v27.8b, v1.8b, v2.8b
> +        uadalp          v16.4s, v29.8h

The scheduling here can be improved, please see the attached patch.

Other than that, it does look reasonable.

// Martin