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

[Martin Storsjö]

On Mon, 22 Aug 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         | 97 ++++++++++++++++++++++++
> 2 files changed, 101 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 d4c0099854..279bae7cb5 100644
> --- a/libavcodec/aarch64/me_cmp_neon.S
> +++ b/libavcodec/aarch64/me_cmp_neon.S
> @@ -659,3 +659,100 @@ 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
> +
> +        movi            v30.4s, #0
> +        movi            v29.4s, #0
> +
> +        add             x5, x1, x3                      // pix1 + stride
> +        add             x6, x2, x3                      // pix2 + stride
> +        sub             w4, w4, #1                      // we need to make h-1 iterations
> +        cmp             w4, #3                          // check if we can make 4 iterations at once
> +        b.le            2f
> +
> +// make 4 iterations at once

The comments seem to talk about 4 iterations at once while the code 
actually only does 3.

> +1:
> +        // x = abs(pix1[0] - pix2[0] - pix1[0 + stride] + pix2[0 + stride]) =

The comment seems a bit un-updated here, since there's no abs() involved 
here

> +        // res = (x) * (x)
> +        ld1             {v0.16b}, [x1], x3              // Load pix1[0], first iteration
> +        ld1             {v1.16b}, [x2], x3              // Load pix2[0], first iteration
> +        ld1             {v2.16b}, [x5], x3              // Load pix1[0 + stride], first iteration
> +        usubl           v28.8h, v0.8b, v1.8b            // Signed difference of pix1[0] - pix2[0], first iteration
> +        ld1             {v3.16b}, [x6], x3              // Load pix2[0 + stride], first iteration
> +        usubl2          v27.8h, v0.16b, v1.16b          // Signed difference of pix1[0] - pix2[0], first iteration
> +        usubl           v26.8h, v3.8b, v2.8b            // Signed difference of pix1[0 + stride] - pix2[0 + stride], first iteration
> +        usubl2          v25.8h, v3.16b, v2.16b          // Signed difference of pix1[0 + stride] - pix2[0 + stride], first iteration

Same thing about reusing data from the previous row, as for the previous 
patch.

// Martin