[FFmpeg-devel] [PATCH 4/4] lavc/vvc_mc R-V V sad

[Rémi Denis-Courmont]

Le sunnuntaina 17. marraskuuta 2024, 15.16.23 EET uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> 
>                             k230               banana_f3
> sad_8x16_c:                  385.9 ( 1.00x)    403.1 ( 1.00x)
> sad_8x16_rvv_i32:            108.1 ( 3.57x)    100.8 ( 4.00x)
> sad_16x8_c:                  376.6 ( 1.00x)    392.6 ( 1.00x)
> sad_16x8_rvv_i32:             89.3 ( 4.21x)    69.5 ( 5.64x)
> sad_16x16_c:                 746.6 ( 1.00x)    757.3 ( 1.00x)
> sad_16x16_rvv_i32:           135.8 ( 5.50x)    121.5 ( 6.23x)
> ---
>  libavcodec/riscv/vvc/Makefile      |  3 +-
>  libavcodec/riscv/vvc/vvc_sad_rvv.S | 58 ++++++++++++++++++++++++++++++
>  libavcodec/riscv/vvc/vvcdsp_init.c |  7 ++++
>  3 files changed, 67 insertions(+), 1 deletion(-)
>  create mode 100644 libavcodec/riscv/vvc/vvc_sad_rvv.S
> 
> diff --git a/libavcodec/riscv/vvc/Makefile b/libavcodec/riscv/vvc/Makefile
> index 582b051579..6b9c618b33 100644
> --- a/libavcodec/riscv/vvc/Makefile
> +++ b/libavcodec/riscv/vvc/Makefile
> @@ -1,2 +1,3 @@
>  OBJS-$(CONFIG_VVC_DECODER) += riscv/vvc/vvcdsp_init.o
> -RVV-OBJS-$(CONFIG_VVC_DECODER) += riscv/vvc/vvc_mc_rvv.o
> +RVV-OBJS-$(CONFIG_VVC_DECODER) += riscv/vvc/vvc_mc_rvv.o \
> +                                  riscv/vvc/vvc_sad_rvv.o
> diff --git a/libavcodec/riscv/vvc/vvc_sad_rvv.S
> b/libavcodec/riscv/vvc/vvc_sad_rvv.S new file mode 100644
> index 0000000000..acdc78d20d
> --- /dev/null
> +++ b/libavcodec/riscv/vvc/vvc_sad_rvv.S
> @@ -0,0 +1,58 @@
> +/*
> + * Copyright (c) 2024 Institue of Software Chinese Academy of Sciences
> (ISCAS). + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA + */
> +
> +#include "libavcodec/riscv/h26x/asm.S"
> +
> +.macro func_sad vlen
> +func ff_vvc_sad_rvv_\vlen, zve32x, zbb, zba
> +        lpad    0
> +        slli              t2, a3, 7  // dy * 128
> +        li                t1, 4*128+4
> +        add               t0, t2, a2 // dy * 128 + dx
> +        sub               t1, t1, t2
> +        sub               t1, t1, a2
> +        sh1add            a0, t0, a0
> +        sh1add            a1, t1, a1
> +        vsetvlstatic32    1, \vlen
> +        li                t0, 16
> +        vmv.s.x           v0, zero
> +        beq               a4, t0, SAD\vlen\()16
> +        .irp w,8,16
> +SAD\vlen\w:
> +        vsetvlstatic16    \w, \vlen
> +        addi              a5, a5, -2
> +        vle16.v           v8, (a0)
> +        vle16.v           v16, (a1)
> +        vwsub.vv          v24, v8, v16

It might be faster to stick to 16-bit and only widen when accumulating. The 
distance between two 16-bit values is an unsigned 16-bit value afterall.

> +        vsetvlstatic32    \w, \vlen
> +        vneg.v            v16, v24
> +        addi              a0, a0, 2 * 128 * 2
> +        vmax.vv           v24, v24, v16
> +        vredsum.vs        v0, v24, v0

Don't calculate a reduction in a loop.

> +        addi              a1, a1, 2 * 128 * 2
> +        bnez              a5, SAD\vlen\w
> +        vmv.x.s           a0, v0
> +        ret
> +        .endr
> +endfunc
> +.endm
> +
> +func_sad 256
> +func_sad 128
> diff --git a/libavcodec/riscv/vvc/vvcdsp_init.c
> b/libavcodec/riscv/vvc/vvcdsp_init.c index 2fe93029aa..1b228cc9f5 100644
> --- a/libavcodec/riscv/vvc/vvcdsp_init.c
> +++ b/libavcodec/riscv/vvc/vvcdsp_init.c
> @@ -59,6 +59,9 @@ DMVR_PROTOTYPES(8, rvv_256)
>      c->inter.dmvr[1][1]   = ff_vvc_dmvr_hv_##bd##_##opt;           \
>  } while (0)
> 
> +int ff_vvc_sad_rvv_128(const int16_t *src0, const int16_t *src1, int dx,
> int dy, int block_w, int block_h); +int ff_vvc_sad_rvv_256(const int16_t
> *src0, const int16_t *src1, int dx, int dy, int block_w, int block_h); +
>  #define PUT_PIXELS_PROTOTYPES2(bd, opt)                                    
>      \ void bf(ff_vvc_put_pixels, bd, opt)(int16_t *dst,                   
>             \ const uint8_t *_src, const ptrdiff_t _src_stride,            
>                \ @@ -97,6 +100,8 @@ void
> ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const int bd) FUNCS(LUMA,
> rvv_256);
>                  FUNCS(CHROMA, rvv_256);
>                  break;
> +            case 10:
> +                c->inter.sad      = ff_vvc_sad_rvv_256;
>              default:
>                  break;
>          }
> @@ -111,6 +116,8 @@ void ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const
> int bd) FUNCS(LUMA, rvv_128);
>                  FUNCS(CHROMA, rvv_128);
>                  break;
> +            case 10:
> +                c->inter.sad      = ff_vvc_sad_rvv_128;
>              default:
>                  break;
>          }


-- 
雷米‧德尼-库尔蒙
http://www.remlab.net/