[FFmpeg-devel] [PATCH 3/5] lavc/aarch64: Add neon implementation for pix_abs16_y2

[Martin Storsjö]

On Tue, 16 Aug 2022, Hubert Mazur wrote:

> Provide optimized implementation of pix_abs16_y2 function for arm64.
>
> Performance comparison tests are shown below.
> pix_abs_0_2_c: 317.2
> pix_abs_0_2_neon: 37.5
>
> 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 |  3 +
> libavcodec/aarch64/me_cmp_neon.S         | 75 ++++++++++++++++++++++++
> 2 files changed, 78 insertions(+)
>
> diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
> index 955592625a..1c36d3d7cb 100644
> --- a/libavcodec/aarch64/me_cmp_init_aarch64.c
> +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
> @@ -29,6 +29,8 @@ int ff_pix_abs16_xy2_neon(MpegEncContext *s, const uint8_t *blk1, const uint8_t
>                       ptrdiff_t stride, int h);
> int ff_pix_abs16_x2_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
>                       ptrdiff_t stride, int h);
> +int ff_pix_abs16_y2_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
> +                      ptrdiff_t stride, int h);

Misaligned function declaration.

> diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
> index 367924b3c2..0ec9c0465b 100644
> --- a/libavcodec/aarch64/me_cmp_neon.S
> +++ b/libavcodec/aarch64/me_cmp_neon.S
> @@ -404,3 +404,78 @@ function sse4_neon, export=1
>
>         ret
> endfunc
> +
> +function ff_pix_abs16_y2_neon, export=1

Why place this new function at the bottom of the file, instead of 
logically following the other preexisting pix_abs16 function? In the 
version I pushed, I moved it further up

> +        // x0           unused
> +        // x1           uint8_t *pix1
> +        // x2           uint8_t *pix2
> +        // x3           ptrdiff_t stride
> +        // x4           int h

This should be w4. You had fixed this in a couple patches, but missed this 
one.

> +
> +        // initialize buffers
> +        movi            v29.8h, #0                      // clear the accumulator
> +        movi            v28.8h, #0                      // clear the accumulator
> +        movi            d18, #0

Unused d18 here too


> +        add             x5, x2, x3                      // pix2 + stride
> +        cmp             w4, #4
> +        b.lt            2f
> +
> +// make 4 iterations at once
> +1:
> +
> +        // abs(pix1[0], avg2(pix2[0], pix2[0 + stride]))
> +        // avg2(a, b) = (((a) + (b) + 1) >> 1)
> +        // abs(x) = (x < 0 ? (-x) : (x))
> +
> +        ld1             {v1.16b}, [x2], x3              // Load pix2 for first iteration
> +        ld1             {v2.16b}, [x5], x3              // Load pix3 for first iteration
> +        urhadd          v30.16b, v1.16b, v2.16b         // Rounding halving add, first iteration
> +        ld1             {v0.16b}, [x1], x3              // Load pix1 for first iteration
> +        uabal           v29.8h, v0.8b, v30.8b           // Absolute difference of lower half, first iteration

This whole first sequence is almost entirely blocking, waiting for the 
result of the previous operation - did you miss to interleave this with 
the rest of the operations?

Normally I wouldn't bother with minor interleaving details, but here the 
impact was rather big. I manually reinterleaved the whole function, and 
got this speedup:

Before:       Cortex A53    A72     A73
pix_abs_0_2_neon:  153.0   63.7    52.7
After:
pix_abs_0_2_neon:  141.0   61.7    51.7

// Martin