[FFmpeg-devel] [PATCH 1/5] lavc/vvc_mc: R-V V put_pixels

Rémi Denis-Courmont remi at remlab.net
Sat Oct 19 13:52:22 EEST 2024


Le perjantaina 11. lokakuuta 2024, 13.38.42 EEST uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> 
>                                                       k230              
> banana_f3 put_chroma_pixels_8_4x4_c:                              61.5 (
> 1.00x)    69.5 ( 1.00x) put_chroma_pixels_8_4x4_rvv_i32:                   
>     33.8 ( 1.82x)    38.2 ( 1.82x) put_chroma_pixels_8_8x8_c:              
>               219.1 ( 1.00x)    246.5 ( 1.00x)
> put_chroma_pixels_8_8x8_rvv_i32:                        70.8 ( 3.09x)   
> 59.2 ( 4.16x) put_chroma_pixels_8_16x16_c:                           830.3
> ( 1.00x)    830.0 ( 1.00x) put_chroma_pixels_8_16x16_rvv_i32:              
>       181.8 ( 4.57x)    142.5 ( 5.82x) put_chroma_pixels_8_32x32_c:        
>                  3246.8 ( 1.00x)    3247.0 ( 1.00x)
> put_chroma_pixels_8_32x32_rvv_i32:                     506.1 ( 6.42x)   
> 330.0 ( 9.84x) put_chroma_pixels_8_64x64_c:                         13616.8
> ( 1.00x)    12967.0 ( 1.00x) put_chroma_pixels_8_64x64_rvv_i32:            
>        1672.5 ( 8.14x)    996.8 (13.01x) put_chroma_pixels_8_128x128_c:    
>                   54263.6 ( 1.00x)    52002.2 ( 1.00x)
> put_chroma_pixels_8_128x128_rvv_i32:                  9181.8 ( 5.91x)   
> 3611.8 (14.40x) put_luma_pixels_8_4x4_c:                               
> 61.5 ( 1.00x)    69.8 ( 1.00x) put_luma_pixels_8_4x4_rvv_i32:              
>            33.8 ( 1.82x)    28.0 ( 2.49x) put_luma_pixels_8_8x8_c:         
>                      219.1 ( 1.00x)    226.0 ( 1.00x)
> put_luma_pixels_8_8x8_rvv_i32:                          80.3 ( 2.73x)   
> 59.2 ( 3.81x) put_luma_pixels_8_16x16_c:                             830.3
> ( 1.00x)    830.2 ( 1.00x) put_luma_pixels_8_16x16_rvv_i32:                
>       181.8 ( 4.57x)    142.5 ( 5.83x) put_luma_pixels_8_32x32_c:          
>                  3256.1 ( 1.00x)    3247.2 ( 1.00x)
> put_luma_pixels_8_32x32_rvv_i32:                       506.1 ( 6.43x)   
> 361.2 ( 8.99x) put_luma_pixels_8_64x64_c:                           14042.5
> ( 1.00x)    12977.2 ( 1.00x) put_luma_pixels_8_64x64_rvv_i32:              
>        1681.8 ( 8.35x)    1007.0 (12.89x) put_luma_pixels_8_128x128_c:     
>                    54291.3 ( 1.00x)    53700.5 ( 1.00x)
> put_luma_pixels_8_128x128_rvv_i32:                   10700.0 ( 5.07x)   
> 3684.8 (14.57x) ---
>  libavcodec/riscv/vvc/vvc_mc_rvv.S  | 61 ++++++++++++++++++++++++++----
>  libavcodec/riscv/vvc/vvcdsp_init.c | 22 +++++++++++
>  2 files changed, 76 insertions(+), 7 deletions(-)
> 
> diff --git a/libavcodec/riscv/vvc/vvc_mc_rvv.S
> b/libavcodec/riscv/vvc/vvc_mc_rvv.S index 2c634af48f..7300889430 100644
> --- a/libavcodec/riscv/vvc/vvc_mc_rvv.S
> +++ b/libavcodec/riscv/vvc/vvc_mc_rvv.S
> @@ -158,7 +158,7 @@
>  .endm
> 
> 
> -.macro AVG_JMP_TABLE id, vlen
> +.macro POW2_JMP_TABLE id, vlen
>  const jmp_table_\id\vlen
>          .4byte \id\()2\vlen\()f - jmp_table_\id\vlen
>          .4byte \id\()4\vlen\()f - jmp_table_\id\vlen
> @@ -170,8 +170,8 @@ const jmp_table_\id\vlen
>  endconst
>  .endm
> 
> -.macro AVG_J vlen, id
> -        clz               t1, a4
> +.macro POW2_J vlen, id, w
> +        clz               t1, \w
>          neg               t1, t1
>          lla               t5, jmp_table_\id\vlen
>          sh2add            t1, t1, t5
> @@ -183,9 +183,9 @@ endconst
>  .macro func_avg vlen
>  func ff_vvc_avg_8_rvv_\vlen\(), zve32x, zbb, zba
>          lpad    0
> -        AVG_JMP_TABLE     1, \vlen
> +        POW2_JMP_TABLE    1, \vlen
>          csrwi             vxrm, 0
> -        AVG_J             \vlen, 1
> +        POW2_J            \vlen, 1, a4
>          .irp w,2,4,8,16,32,64,128
>          avg               \w, \vlen, 1
>          .endr
> @@ -265,7 +265,7 @@ func_avg 256
>  .macro func_w_avg vlen
>  func ff_vvc_w_avg_8_rvv_\vlen\(), zve32x, zbb, zba
>          lpad    0
> -        AVG_JMP_TABLE     2, \vlen
> +        POW2_JMP_TABLE    2, \vlen
>          csrwi             vxrm, 0
>          addi              t6, a6, 7
>          ld                t3, (sp)
> @@ -275,7 +275,7 @@ func ff_vvc_w_avg_8_rvv_\vlen\(), zve32x, zbb, zba
>          add               t4, t4, t5
>          addi              t5, t6, -1      // shift - 1
>          sll               t4, t4, t5
> -        AVG_J             \vlen, 2
> +        POW2_J            \vlen, 2, a4
>          .irp w,2,4,8,16,32,64,128
>          w_avg             \w, \vlen, 2
>          .endr
> @@ -405,3 +405,50 @@ func_dmvr \vlen, dmvr_h
>  func_dmvr \vlen, dmvr_v
>  func_dmvr \vlen, dmvr_hv
>  .endr
> +
> +.macro put_pixels w, vlen, id
> +\id\w\vlen:
> +        vsetvlstatic16    \w, \vlen
> +.if \w == 128 && \vlen == 128
> +1:
> +        addi              t0, a1, 64
> +        addi              t1, a0, 64*2
> +        vle8.v            v0, (a1)
> +        vle8.v            v16, (t0)
> +        vzext.vf2         v8, v0
> +        vzext.vf2         v24, v16
> +        vsll.vi           v8, v8, 6
> +        vsll.vi           v24, v24, 6

Even without Zvbb's widening shift, widening multiplication is probably faster 
here.

> +        vse16.v           v8, (a0)
> +        vse16.v           v24, (t1)
> +        add               a1, a1, a2
> +        addi              a3, a3, -1
> +        addi              a0, a0, 128*2
> +        bnez              a3, 1b
> +.else
> +1:
> +        vle8.v            v0, (a1)
> +        vzext.vf2         v8, v0
> +        vsll.vi           v8, v8, 6
> +        vse16.v           v8, (a0)
> +        add               a1, a1, a2
> +        addi              a3, a3, -1
> +        addi              a0, a0, 128*2
> +        bnez              a3, 1b
> +.endif
> +        ret
> +.endm
> +
> +.macro func_put_pixels vlen
> +func ff_vvc_put_pixels_8_rvv_\vlen\(), zve32x, zbb, zba
> +        lpad    0
> +        POW2_JMP_TABLE    3, \vlen
> +        POW2_J            \vlen, 3, a6
> +        .irp w,2,4,8,16,32,64,128
> +        put_pixels        \w, \vlen, 3
> +        .endr
> +endfunc
> +.endm
> +
> +func_put_pixels 256
> +func_put_pixels 128
> diff --git a/libavcodec/riscv/vvc/vvcdsp_init.c
> b/libavcodec/riscv/vvc/vvcdsp_init.c index 7df3ce58db..18e7a455f2 100644
> --- a/libavcodec/riscv/vvc/vvcdsp_init.c
> +++ b/libavcodec/riscv/vvc/vvcdsp_init.c
> @@ -24,6 +24,7 @@
>  #include "libavutil/cpu.h"
>  #include "libavutil/riscv/cpu.h"
>  #include "libavcodec/vvc/dsp.h"
> +#include "libavcodec/vvc/dec.h"
> 
>  #define bf(fn, bd,  opt) fn##_##bd##_##opt
> 
> @@ -57,6 +58,23 @@ DMVR_PROTOTYPES(8, rvv_256)
>      c->inter.dmvr[1][1]   = ff_vvc_dmvr_hv_##bd##_##opt;           \
>  } while (0)
> 
> +#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,      
>                      \ +    const int height, const int8_t *hf, const
> int8_t *vf, const int width); +
> +PUT_PIXELS_PROTOTYPES2(8, rvv_128)
> +PUT_PIXELS_PROTOTYPES2(8, rvv_256)
> +
> +#define PEL_FUNC(dst, C, idx1, idx2, a)                                    
>       \ +    do {                                                          
>                \ +        for (int w = 1; w < 7; w++)                      
>                         \ +            c->inter.dst[C][w][idx1][idx2] = a; 
>                                  \ +    } while (0)                        
>                                           \ +
> +#define FUNCS(C, opt)                                                      
>       \ +        PEL_FUNC(put, C, 0, 0, ff_vvc_put_pixels_8_##opt);        
>                \ +
>  void ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const int bd)
>  {
>  #if HAVE_RVV
> @@ -72,6 +90,8 @@ void ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const
> int bd) c->inter.w_avg    = ff_vvc_w_avg_8_rvv_256;
>  # endif
>                  DMVR_INIT(8, rvv_256);
> +                FUNCS(LUMA, rvv_256);
> +                FUNCS(CHROMA, rvv_256);
>                  break;
>              default:
>                  break;
> @@ -85,6 +105,8 @@ void ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const
> int bd) c->inter.w_avg    = ff_vvc_w_avg_8_rvv_128;
>  # endif
>                  DMVR_INIT(8, rvv_128);
> +                FUNCS(LUMA, rvv_128);
> +                FUNCS(CHROMA, rvv_128);
>                  break;
>              default:
>                  break;


-- 
レミ・デニ-クールモン
http://www.remlab.net/





More information about the ffmpeg-devel mailing list