[FFmpeg-devel] [PATCH 2/5] lavc/aarch64: Add neon implementation of vsse16

[Martin Storsjö]

On Tue, 6 Sep 2022, Hubert Mazur wrote:

> Provide optimized implementation of vsse16 for arm64.
>
> Performance comparison tests are shown below.
> - vsse_0_c: 254.4
> - vsse_0_neon: 64.7
>
> Benchmarks and tests are run with checkasm tool on AWS Graviton 3.
>
> Signed-off-by: Hubert Mazur <hum at semihalf.com>
> ---
> libavcodec/aarch64/me_cmp_init_aarch64.c |  4 ++
> libavcodec/aarch64/me_cmp_neon.S         | 87 ++++++++++++++++++++++++
> 2 files changed, 91 insertions(+)
>
> diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
> index ddc5d05611..7b81e48d16 100644
> --- a/libavcodec/aarch64/me_cmp_init_aarch64.c
> +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
> @@ -43,6 +43,8 @@ int sse4_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
> 
> int vsad16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
>                 ptrdiff_t stride, int h);
> +int vsse16_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)
> {
> @@ -62,5 +64,7 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
>         c->sse[2] = sse4_neon;
>
>         c->vsad[0] = vsad16_neon;
> +
> +        c->vsse[0] = vsse16_neon;
>     }
> }
> diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
> index 1d0b166d69..b3f376aa60 100644
> --- a/libavcodec/aarch64/me_cmp_neon.S
> +++ b/libavcodec/aarch64/me_cmp_neon.S
> @@ -649,3 +649,90 @@ function vsad16_neon, export=1
>
>         ret
> endfunc
> +
> +function vsse16_neon, export=1
> +        // x0           unused
> +        // x1           uint8_t *pix1
> +        // x2           uint8_t *pix2
> +        // x3           ptrdiff_t stride
> +        // w4           int h
> +
> +        ld1             {v0.16b}, [x1], x3              // Load pix1[0], first iteration
> +        ld1             {v1.16b}, [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
> +        usubl2          v30.8h, v0.16b, v1.16b          // Signed difference of pix1[0] - pix2[0], first iteration
> +        b.le            2f
> +
> +
> +1:
> +        // x = abs(pix1[0] - pix2[0] - pix1[0 + stride] + pix2[0 + stride])
> +        // res = (x) * (x)

Technically, there's no need for abs() here, we can just as well just do a 
plain subtraction. I tested this by replacing sabd with sub here (and 
changing umlal into smlal). It doesn't make any difference for the 
performance on the cores I tested on though - apparently there's no 
difference in performance between sabd and sub. So in practice, both 
should be fine. And I don't think that either of them is better for 
handling overflows/edge cases here either (which shouldn't be happening 
anyway).

// Martin