[FFmpeg-devel] [PATCH v4 2/4] lavc/vp9dsp: R-V V mc bilin hv

[Rémi Denis-Courmont]

Le lauantaina 15. kesäkuuta 2024, 14.50.32 EEST uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> 
>                                                      C908   X60
> vp9_avg_bilin_4hv_8bpp_c                           :   10.7    9.5
> vp9_avg_bilin_4hv_8bpp_rvv_i32                     :    4.0    3.5
> vp9_avg_bilin_8hv_8bpp_c                           :   38.5   34.2
> vp9_avg_bilin_8hv_8bpp_rvv_i32                     :    7.2    6.5
> vp9_avg_bilin_16hv_8bpp_c                          :  147.2  130.5
> vp9_avg_bilin_16hv_8bpp_rvv_i32                    :   14.5   12.7
> vp9_avg_bilin_32hv_8bpp_c                          :  574.2  509.7
> vp9_avg_bilin_32hv_8bpp_rvv_i32                    :   42.5   38.0
> vp9_avg_bilin_64hv_8bpp_c                          : 2321.2 2017.7
> vp9_avg_bilin_64hv_8bpp_rvv_i32                    :  163.5  131.0
> vp9_put_bilin_4hv_8bpp_c                           :   10.0    8.7
> vp9_put_bilin_4hv_8bpp_rvv_i32                     :    3.5    3.0
> vp9_put_bilin_8hv_8bpp_c                           :   35.2   31.2
> vp9_put_bilin_8hv_8bpp_rvv_i32                     :    6.5    5.7
> vp9_put_bilin_16hv_8bpp_c                          :  134.0  119.0
> vp9_put_bilin_16hv_8bpp_rvv_i32                    :   12.7   11.5
> vp9_put_bilin_32hv_8bpp_c                          :  538.5  464.2
> vp9_put_bilin_32hv_8bpp_rvv_i32                    :   39.7   35.2
> vp9_put_bilin_64hv_8bpp_c                          : 2111.7 1833.2
> vp9_put_bilin_64hv_8bpp_rvv_i32                    :  138.5  122.5
> ---
>  libavcodec/riscv/vp9_mc_rvv.S  | 38 +++++++++++++++++++++++++++++++++-
>  libavcodec/riscv/vp9dsp_init.c | 10 +++++++++
>  2 files changed, 47 insertions(+), 1 deletion(-)
> 
> diff --git a/libavcodec/riscv/vp9_mc_rvv.S b/libavcodec/riscv/vp9_mc_rvv.S
> index fb7377048a..5241562531 100644
> --- a/libavcodec/riscv/vp9_mc_rvv.S
> +++ b/libavcodec/riscv/vp9_mc_rvv.S
> @@ -147,6 +147,40 @@ func ff_\op\()_vp9_bilin_64\type\()_rvv, zve32x
>  endfunc
>  .endm
> 
> +.macro bilin_hv op
> +func ff_\op\()_vp9_bilin_64hv_rvv, zve32x
> +        vsetvlstatic8   64, t0, 64
> +.Lbilin_hv\op:
> +.ifc \op,avg
> +        csrwi           vxrm, 0
> +.endif
> +        neg             t1, a5
> +        neg             t2, a6
> +        li              t4, 8
> +        bilin_load_h    v24, put, a5
> +        add             a2, a2, a3
> +1:
> +        addi            a4, a4, -1
> +        bilin_load_h    v4, put, a5
> +        vwmulu.vx       v16, v4, a6
> +        vwmaccsu.vx     v16, t2, v24
> +        vwadd.wx        v16, v16, t4
> +        vnsra.wi        v16, v16, 4

Why round manually?
It looks like vnclip.wi would be more straightforward here.

> +        vadd.vv         v0, v16, v24
> +.ifc \op,avg
> +        vle8.v          v16, (a0)
> +        vaaddu.vv       v0, v0, v16
> +.endif
> +        vse8.v          v0, (a0)
> +        vmv.v.v         v24, v4
> +        add             a2, a2, a3
> +        add             a0, a0, a1
> +        bnez            a4, 1b
> +
> +        ret
> +endfunc
> +.endm
> +
>  .irp len, 64, 32, 16, 8, 4
>          copy_avg \len
>  .endr
> @@ -155,6 +189,8 @@ bilin_h_v  put, h, a5
>  bilin_h_v  avg, h, a5
>  bilin_h_v  put, v, a6
>  bilin_h_v  avg, v, a6
> +bilin_hv   put
> +bilin_hv   avg
> 
>  .macro func_bilin_h_v len, op, type
>  func ff_\op\()_vp9_bilin_\len\()\type\()_rvv, zve32x
> @@ -165,7 +201,7 @@ endfunc
> 
>  .irp len, 32, 16, 8, 4
>          .irp op, put, avg
> -                .irp type, h, v
> +                .irp type, h, v, hv
>                          func_bilin_h_v \len, \op, \type
>                  .endr
>          .endr
> diff --git a/libavcodec/riscv/vp9dsp_init.c b/libavcodec/riscv/vp9dsp_init.c
> index 9606d8545f..b3700dfb08 100644
> --- a/libavcodec/riscv/vp9dsp_init.c
> +++ b/libavcodec/riscv/vp9dsp_init.c
> @@ -83,6 +83,16 @@ static av_cold void vp9dsp_mc_init_riscv(VP9DSPContext
> *dsp, int bpp) dsp->mc[4][FILTER_BILINEAR ][0][1][0] =
> ff_put_vp9_bilin_4h_rvv; dsp->mc[4][FILTER_BILINEAR ][1][0][1] =
> ff_avg_vp9_bilin_4v_rvv; dsp->mc[4][FILTER_BILINEAR ][1][1][0] =
> ff_avg_vp9_bilin_4h_rvv; +    dsp->mc[0][FILTER_BILINEAR ][0][1][1] =
> ff_put_vp9_bilin_64hv_rvv; +    dsp->mc[0][FILTER_BILINEAR ][1][1][1] =
> ff_avg_vp9_bilin_64hv_rvv; +    dsp->mc[1][FILTER_BILINEAR ][0][1][1] =
> ff_put_vp9_bilin_32hv_rvv; +    dsp->mc[1][FILTER_BILINEAR ][1][1][1] =
> ff_avg_vp9_bilin_32hv_rvv; +    dsp->mc[2][FILTER_BILINEAR ][0][1][1] =
> ff_put_vp9_bilin_16hv_rvv; +    dsp->mc[2][FILTER_BILINEAR ][1][1][1] =
> ff_avg_vp9_bilin_16hv_rvv; +    dsp->mc[3][FILTER_BILINEAR ][0][1][1] =
> ff_put_vp9_bilin_8hv_rvv; +    dsp->mc[3][FILTER_BILINEAR ][1][1][1] =
> ff_avg_vp9_bilin_8hv_rvv; +    dsp->mc[4][FILTER_BILINEAR ][0][1][1] =
> ff_put_vp9_bilin_4hv_rvv; +    dsp->mc[4][FILTER_BILINEAR ][1][1][1] =
> ff_avg_vp9_bilin_4hv_rvv;
> 
>  #undef init_fpel
>      }


-- 
Rémi Denis-Courmont
http://www.remlab.net/