[FFmpeg-devel] [PATCH] lavc/aarch64: hevc_sao reschedule slightly

[Martin Storsjö]

On Wed, 25 May 2022, J. Dekker wrote:

> Signed-off-by: J. Dekker <jdek at itanimul.li>
> ---
> libavcodec/aarch64/hevcdsp_sao_neon.S | 30 +++++++++++++++------------
> 1 file changed, 17 insertions(+), 13 deletions(-)
>
> diff --git a/libavcodec/aarch64/hevcdsp_sao_neon.S b/libavcodec/aarch64/hevcdsp_sao_neon.S
> index efd8112af4..39056d76ee 100644
> --- a/libavcodec/aarch64/hevcdsp_sao_neon.S
> +++ b/libavcodec/aarch64/hevcdsp_sao_neon.S
> @@ -24,6 +24,10 @@
>
> #include "libavutil/aarch64/asm.S"
>
> +#define MAX_PB_SIZE 64
> +#define AV_INPUT_BUFFER_PADDING_SIZE 64
> +#define SAO_STRIDE (2*MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE)
> +
> // void sao_band_filter(uint8_t *_dst, uint8_t *_src,
> //                      ptrdiff_t stride_dst, ptrdiff_t stride_src,
> //                      int16_t *sao_offset_val, int sao_left_class,
> @@ -56,6 +60,7 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
>         // |xDE#xAD|xCA#xFE|xBE#xEF|xFE#xED|....
>         // +----------------------------------->
>         //    i-0     i-1     i-2     i-3
> +        subs            w8, w8,  #8
>         ld1             {v2.8b}, [x1], #8          // dst[x] = av_clip_pixel(src[x] + offset_table[src[x] >> shift]);

For cases like this, it's usually better to place the subs after the ld1, 
because the ld1 will take a couple cycles before the result is available 
(which the next instruction needs).

>         uxtl            v0.8h,  v2.8b              // load src[x]
>         ushr            v2.8h,  v0.8h, #3          // >> BIT_DEPTH - 3
> @@ -66,7 +71,7 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
>         add             v1.8h,  v0.8h, v2.8h       // src[x] + table
>         sqxtun          v4.8b,  v1.8h              // clip + narrow
>         st1             {v4.8b}, [x0], #8          // store
> -        subs            w8, w8,  #8                // done 8 pixels
> +        // done 8 pixels
>         bne             2b
>         subs            w7, w7,  #1                // finished line, prep. new
>         add             x0, x0,  x2                // dst += stride_dst
> @@ -75,12 +80,11 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
>         ret
> endfunc
>
> -// ASSUMES STRIDE_SRC = 192
> .Lsao_edge_pos:
> .word 1 // horizontal
> -.word 192 // vertical
> -.word 192 + 1 // 45 degree
> -.word 192 - 1 // 135 degree
> +.word SAO_STRIDE // vertical
> +.word SAO_STRIDE + 1 // 45 degree
> +.word SAO_STRIDE - 1 // 135 degree
>
> // ff_hevc_sao_edge_filter_16x16_8_neon(char *dst, char *src, ptrdiff stride_dst,
> //                                      int16 *sao_offset_val, int eo, int width, int height)
> @@ -98,7 +102,7 @@ function ff_hevc_sao_edge_filter_16x16_8_neon, export=1
>         uzp2            v1.16b, v3.16b, v3.16b     // sao_offset_val -> upper
>         uzp1            v0.16b, v3.16b, v3.16b     // sao_offset_val -> lower
>         movi            v2.16b, #2
> -        mov             x15, #192
> +        mov             x15, #SAO_STRIDE
>         // strides between end of line and next src/dst
>         sub             x15, x15, x5               // stride_src - width
>         sub             x16, x2, x5                // stride_dst - width
> @@ -108,6 +112,7 @@ function ff_hevc_sao_edge_filter_16x16_8_neon, export=1
>         sub             x12, x11, x4               // src_a (prev) = src - sao_edge_pos
>         add             x13, x11, x4               // src_b (next) = src + sao_edge_pos
> 2:      // process 16 bytes
> +        subs            x14, x14, #16
>         ld1             {v3.16b}, [x11], #16       // load src
>         ld1             {v4.16b}, [x12], #16       // load src_a (prev)
>         ld1             {v5.16b}, [x13], #16       // load src_b (next)

Same thing here, it's better to do the subs after firing off all loads.

> @@ -130,12 +135,12 @@ function ff_hevc_sao_edge_filter_16x16_8_neon, export=1
>         sqxtun          v3.8b, v20.8h
>         sqxtun2         v3.16b, v21.8h
>         st1             {v3.16b}, [x0], #16
> -        subs            x14, x14, #16              // filtered 16 bytes
> +        // filtered 16 bytes
>         b.ne            2b                         // do we have width to filter?
>         // no width to filter, setup next line
> +        subs            w6, w6, #1                 // filtered line
>         add             x11, x11, x15              // stride src to next line
>         add             x0, x0, x16                // stride dst to next line
> -        subs            w6, w6, #1                 // filtered line
>         b.ne            1b                         // do we have lines to process?

This looks good!

>         // no lines to filter
>         ret
> @@ -156,12 +161,12 @@ function ff_hevc_sao_edge_filter_8x8_8_neon, export=1
>         movi            v2.16b, #2
>         add             x16, x0, x2
>         lsl             x2,  x2, #1
> -        mov             x15, #192
> +        mov             x15, #SAO_STRIDE
>         mov             x8,  x1
>         sub             x9,  x1, x4
>         add             x10, x1, x4
> -        lsr             w17, w6, #1
> -1:      ld1             {v3.d}[0], [ x8], x15
> +1:      subs            w6, w6, #2
> +        ld1             {v3.d}[0], [ x8], x15
>         ld1             {v4.d}[0], [ x9], x15
>         ld1             {v5.d}[0], [x10], x15
>         ld1             {v3.d}[1], [ x8], x15

Ditto, move the decrement after the loads.

// Martin