[FFmpeg-devel] [PATCH 3/4] lavc/aarch64: Provide optimized implementation of vsse8 for arm64.

[Martin Storsjö]

On Mon, 26 Sep 2022, Grzegorz Bernacki wrote:

> Provide optimized implementation of vsse8 for arm64.
>
> Performance comparison tests are shown below.
> - vsse_1_c: 141.5
> - vsse_1_neon: 32.5
>
> Benchmarks and tests are run with checkasm tool on AWS Graviton 3.
>
> Signed-off-by: Grzegorz Bernacki <gjb at semihalf.com>
> ---
> libavcodec/aarch64/me_cmp_init_aarch64.c |  5 ++
> libavcodec/aarch64/me_cmp_neon.S         | 70 ++++++++++++++++++++++++
> 2 files changed, 75 insertions(+)
>
> diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
> index 2c61cfcf63..f247372c94 100644
> --- a/libavcodec/aarch64/me_cmp_init_aarch64.c
> +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
> @@ -71,6 +71,9 @@ int nsse8_neon(int multiplier, const uint8_t *s, const uint8_t *s2,
> int nsse8_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
>                        ptrdiff_t stride, int h);
> 
> +int vsse8_neon(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)
> {
>     int cpu_flags = av_get_cpu_flags();
> @@ -96,6 +99,8 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
>         c->vsad[5] = vsad_intra8_neon;
>
>         c->vsse[0] = vsse16_neon;
> +        c->vsse[1] = vsse8_neon;
> +
>         c->vsse[4] = vsse_intra16_neon;
>
>         c->nsse[0] = nsse16_neon_wrapper;
> diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
> index 6f7c7c1690..386d2de0c5 100644
> --- a/libavcodec/aarch64/me_cmp_neon.S
> +++ b/libavcodec/aarch64/me_cmp_neon.S
> @@ -843,6 +843,76 @@ function vsad16_neon, export=1
>         ret
> endfunc
> 
> +function vsse8_neon, export=1
> +        // x0           unused
> +        // x1           uint8_t *pix1
> +        // x2           uint8_t *pix2
> +        // x3           ptrdiff_t stride
> +        // w4           int h
> +
> +        ld1             {v0.8b}, [x1], x3              // Load pix1[0], first iteration
> +        ld1             {v1.8b}, [x2], x3              // Load pix2[0], first iteration
> +
> +        sub             w4, w4, #1                      // we need to make h-1 iterations
> +        movi            v16.4s, #0
> +        movi            v17.4s, #0
> +
> +        cmp             w4, #3                          // check if we can make 3 iterations at once
> +        usubl           v31.8h, v0.8b, v1.8b            // Signed difference of pix1[0] - pix2[0], first iteration
> +        b.le            2f

Why the b.le here, shouldn't it be enough with b.lt? If we're run with 
h=<unroll amount>, it should be enough to run one round in the unrolled 
loop.

I see that we've got the same issue in a couple preexisting functions; 
I'll send a patch to fix that.

Other than that, this looks reasonable to me.

// Martin