[FFmpeg-devel] [PATCH 1/3] lavc/vvc_mc: R-V V put_pixels

Rémi Denis-Courmont remi at remlab.net
Sat Nov 16 10:37:07 EET 2024


Le sunnuntaina 10. marraskuuta 2024, 13.34.31 EET uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> 
>                                                       k230              
> banana_f3 put_chroma_pixels_8_4x4_c:                              63.5 (
> 1.00x)    59.2 ( 1.00x) put_chroma_pixels_8_4x4_rvv_i32:                   
>     26.5 ( 2.39x)    28.0 ( 2.12x) put_chroma_pixels_8_8x8_c:              
>               211.8 ( 1.00x)    215.5 ( 1.00x)
> put_chroma_pixels_8_8x8_rvv_i32:                        54.3 ( 3.90x)   
> 48.8 ( 4.42x) put_chroma_pixels_8_16x16_c:                           841.3
> ( 1.00x)    830.0 ( 1.00x) put_chroma_pixels_8_16x16_rvv_i32:              
>       137.5 ( 6.12x)    121.8 ( 6.82x) put_chroma_pixels_8_32x32_c:        
>                  3248.8 ( 1.00x)    3288.2 ( 1.00x)
> put_chroma_pixels_8_32x32_rvv_i32:                     350.5 ( 9.27x)   
> 288.5 (11.40x) put_chroma_pixels_8_64x64_c:                         12998.3
> ( 1.00x)    12976.2 ( 1.00x) put_chroma_pixels_8_64x64_rvv_i32:            
>        1100.5 (11.81x)    924.0 (14.04x) put_chroma_pixels_8_128x128_c:    
>                   54284.0 ( 1.00x)    52654.5 ( 1.00x)
> put_chroma_pixels_8_128x128_rvv_i32:                  7192.8 ( 7.55x)   
> 2934.2 (17.94x) put_luma_pixels_8_4x4_c:                               
> 63.5 ( 1.00x)    69.5 ( 1.00x) put_luma_pixels_8_4x4_rvv_i32:              
>            26.5 ( 2.39x)    28.0 ( 2.48x) put_luma_pixels_8_8x8_c:         
>                      211.5 ( 1.00x)    225.8 ( 1.00x)
> put_luma_pixels_8_8x8_rvv_i32:                          54.3 ( 3.90x)   
> 38.5 ( 5.86x) put_luma_pixels_8_16x16_c:                             850.5
> ( 1.00x)    830.0 ( 1.00x) put_luma_pixels_8_16x16_rvv_i32:                
>       137.5 ( 6.18x)    100.8 ( 8.24x) put_luma_pixels_8_32x32_c:          
>                  3248.8 ( 1.00x)    3257.2 ( 1.00x)
> put_luma_pixels_8_32x32_rvv_i32:                       341.3 ( 9.52x)   
> 246.8 (13.20x) put_luma_pixels_8_64x64_c:                           13007.5
> ( 1.00x)    13038.8 ( 1.00x) put_luma_pixels_8_64x64_rvv_i32:              
>        1119.0 (11.62x)    684.2 (19.06x) put_luma_pixels_8_128x128_c:      
>                   54219.3 ( 1.00x)    52060.8 ( 1.00x)
> put_luma_pixels_8_128x128_rvv_i32:                    6813.5 ( 7.96x)   
> 2548.8 (20.43x) ---
>  libavcodec/riscv/h26x/asm.S        | 169 +++++++++++++++++++++++++++++
>  libavcodec/riscv/h26x/h2656dsp.h   |  27 +++++
>  libavcodec/riscv/vvc/vvc_mc_rvv.S  | 120 ++------------------
>  libavcodec/riscv/vvc/vvcdsp_init.c |  23 ++++
>  4 files changed, 227 insertions(+), 112 deletions(-)
>  create mode 100644 libavcodec/riscv/h26x/asm.S
>  create mode 100644 libavcodec/riscv/h26x/h2656dsp.h
> 
> diff --git a/libavcodec/riscv/h26x/asm.S b/libavcodec/riscv/h26x/asm.S
> new file mode 100644
> index 0000000000..d37b459f66
> --- /dev/null
> +++ b/libavcodec/riscv/h26x/asm.S
> @@ -0,0 +1,169 @@
> +/*
> + * 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 "libavutil/riscv/asm.S"
> +
> +.macro vsetvlstatic8 w, vlen
> +        .if \w == 2 && \vlen == 128
> +                vsetivli        zero, \w, e8, mf8, ta, ma
> +        .elseif \w <= 4 && \vlen == 128
> +                vsetivli        zero, \w, e8, mf4, ta, ma
> +        .elseif \w <= 8 && \vlen == 128
> +                vsetivli        zero, \w, e8, mf2, ta, ma
> +        .elseif \w <= 16 && \vlen == 128
> +                vsetivli        zero, \w, e8, m1, ta, ma
> +        .elseif \w <= 32 && \vlen == 128
> +                li              t0, \w
> +                vsetvli         zero, t0, e8, m2, ta, ma
> +        .elseif \w <= 4 && \vlen == 256
> +                vsetivli        zero, \w, e8, mf8, ta, ma
> +        .elseif \w <= 8 && \vlen == 256
> +                vsetivli        zero, \w, e8, mf4, ta, ma
> +        .elseif \w <= 16 && \vlen == 256
> +                vsetivli        zero, \w, e8, mf2, ta, ma
> +        .elseif \w <= 32 && \vlen == 256
> +                li              t0, \w
> +                vsetvli         zero, t0, e8, m1, ta, ma
> +        .elseif \w <= 64 && \vlen == 256
> +                li              t0, \w
> +                vsetvli         zero, t0, e8, m2, ta, ma
> +        .else
> +                li              t0, \w
> +                vsetvli         zero, t0, e8, m4, ta, ma
> +        .endif
> +.endm
> +
> +.macro vsetvlstatic16 w, vlen
> +        .if \w == 2 && \vlen == 128
> +                vsetivli        zero, \w, e16, mf4, ta, ma
> +        .elseif \w <= 4 && \vlen == 128
> +                vsetivli        zero, \w, e16, mf2, ta, ma
> +        .elseif \w <= 8 && \vlen == 128
> +                vsetivli        zero, \w, e16, m1, ta, ma
> +        .elseif \w <= 16 && \vlen == 128
> +                vsetivli        zero, \w, e16, m2, ta, ma
> +        .elseif \w <= 32 && \vlen == 128
> +                li              t0, \w
> +                vsetvli         zero, t0, e16, m4, ta, ma
> +        .elseif \w <= 4 && \vlen == 256
> +                vsetivli        zero, \w, e16, mf4, ta, ma
> +        .elseif \w <= 8 && \vlen == 256
> +                vsetivli        zero, \w, e16, mf2, ta, ma
> +        .elseif \w <= 16 && \vlen == 256
> +                vsetivli        zero, \w, e16, m1, ta, ma
> +        .elseif \w <= 32 && \vlen == 256
> +                li              t0, \w
> +                vsetvli         zero, t0, e16, m2, ta, ma
> +        .elseif \w <= 64 && \vlen == 256
> +                li              t0, \w
> +                vsetvli         zero, t0, e16, m4, ta, ma
> +        .else
> +                li              t0, \w
> +                vsetvli         zero, t0, e16, m8, ta, ma
> +        .endif
> +.endm
> +
> +.macro vsetvlstatic32 w, vlen
> +        .if \w == 2
> +                vsetivli        zero, \w, e32, mf2, ta, ma
> +        .elseif \w <= 4 && \vlen == 128
> +                vsetivli        zero, \w, e32, m1, ta, ma
> +        .elseif \w <= 8 && \vlen == 128
> +                vsetivli        zero, \w, e32, m2, ta, ma
> +        .elseif \w <= 16 && \vlen == 128
> +                vsetivli        zero, \w, e32, m4, ta, ma
> +        .elseif \w <= 4 && \vlen == 256
> +                vsetivli        zero, \w, e32, mf2, ta, ma
> +        .elseif \w <= 8 && \vlen == 256
> +                vsetivli        zero, \w, e32, m1, ta, ma
> +        .elseif \w <= 16 && \vlen == 256
> +                vsetivli        zero, \w, e32, m2, ta, ma
> +        .elseif \w <= 32 && \vlen == 256
> +                li              t0, \w
> +                vsetvli         zero, t0, e32, m4, ta, ma
> +        .else
> +                li              t0, \w
> +                vsetvli         zero, t0, e32, m8, ta, ma
> +        .endif
> +.endm

Generally speaking, I think that moving code should be done in dedicated 
patches.

> +
> +.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
> +        .4byte \id\()8\vlen\()f - jmp_table_\id\vlen
> +        .4byte \id\()16\vlen\()f - jmp_table_\id\vlen
> +        .4byte \id\()32\vlen\()f - jmp_table_\id\vlen
> +        .4byte \id\()64\vlen\()f - jmp_table_\id\vlen
> +        .4byte \id\()128\vlen\()f - jmp_table_\id\vlen
> +endconst
> +.endm
> +
> +.macro POW2_J vlen, id, w
> +        clz               t1, \w
> +        neg               t1, t1
> +        lla               t5, jmp_table_\id\vlen
> +        sh2add            t1, t1, t5
> +        lw                t1, ((__riscv_xlen-2)<<2)(t1)
> +        add               t1, t1, t5
> +        jr                t1
> +.endm
> +
> +.macro put_pixels w, vlen, id, MAX_PB_SIZE
> +\id\w\vlen:
> +        vsetvlstatic8     \w, \vlen

You can branch here. The rest of the byte code is the same in all but one 
cases.

> +        li                t2, 1<<6
> +.if \w == 128 && \vlen == 128
> +1:
> +        addi              t0, a1, 64
> +        addi              t1, a0, 64*2
> +        vle8.v            v0, (a1)
> +        vle8.v            v16, (t0)
> +        vwmulu.vx         v8, v0, t2
> +        vwmulu.vx         v24, v16, t2
> +        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)
> +        vwmulu.vx         v8, v0, t2
> +        vse16.v           v8, (a0)
> +        add               a1, a1, a2
> +        addi              a3, a3, -1
> +        addi              a0, a0, \MAX_PB_SIZE<<1
> +        bnez              a3, 1b
> +.endif
> +        ret
> +.endm
> +
> +.macro func_put_pixels vlen, MAX_PB_SIZE, name
> +func ff_\name\()_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, \MAX_PB_SIZE
> +        .endr
> +endfunc
> +.endm
> diff --git a/libavcodec/riscv/h26x/h2656dsp.h
> b/libavcodec/riscv/h26x/h2656dsp.h new file mode 100644
> index 0000000000..5ddfb99881
> --- /dev/null
> +++ b/libavcodec/riscv/h26x/h2656dsp.h
> @@ -0,0 +1,27 @@
> +/*
> + * 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 + */
> +
> +#ifndef AVCODEC_RISCV_H26X_H2656DSP_H
> +#define AVCODEC_RISCV_H26X_H2656DSP_H
> +
> +void ff_h2656_put_pixels_8_rvv_256(int16_t *dst, const uint8_t *src,
> ptrdiff_t srcstride, int height, intptr_t mx, intptr_t my, int width);
> +void ff_h2656_put_pixels_8_rvv_128(int16_t *dst, const uint8_t *src,
> ptrdiff_t srcstride, int height, intptr_t mx, intptr_t my, int width); +
> +#endif
> diff --git a/libavcodec/riscv/vvc/vvc_mc_rvv.S
> b/libavcodec/riscv/vvc/vvc_mc_rvv.S index 1dcbaf7d5b..329fc0648c 100644
> --- a/libavcodec/riscv/vvc/vvc_mc_rvv.S
> +++ b/libavcodec/riscv/vvc/vvc_mc_rvv.S
> @@ -18,91 +18,7 @@
>   * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA */
> 
> -#include "libavutil/riscv/asm.S"
> -
> -.macro vsetvlstatic8 w, vlen
> -        .if \w == 2 && \vlen == 128
> -                vsetivli        zero, \w, e8, mf8, ta, ma
> -        .elseif \w <= 4 && \vlen == 128
> -                vsetivli        zero, \w, e8, mf4, ta, ma
> -        .elseif \w <= 8 && \vlen == 128
> -                vsetivli        zero, \w, e8, mf2, ta, ma
> -        .elseif \w <= 16 && \vlen == 128
> -                vsetivli        zero, \w, e8, m1, ta, ma
> -        .elseif \w <= 32 && \vlen == 128
> -                li              t0, \w
> -                vsetvli         zero, t0, e8, m2, ta, ma
> -        .elseif \w <= 4 && \vlen == 256
> -                vsetivli        zero, \w, e8, mf8, ta, ma
> -        .elseif \w <= 8 && \vlen == 256
> -                vsetivli        zero, \w, e8, mf4, ta, ma
> -        .elseif \w <= 16 && \vlen == 256
> -                vsetivli        zero, \w, e8, mf2, ta, ma
> -        .elseif \w <= 32 && \vlen == 256
> -                li              t0, \w
> -                vsetvli         zero, t0, e8, m1, ta, ma
> -        .elseif \w <= 64 && \vlen == 256
> -                li              t0, \w
> -                vsetvli         zero, t0, e8, m2, ta, ma
> -        .else
> -                li              t0, \w
> -                vsetvli         zero, t0, e8, m4, ta, ma
> -        .endif
> -.endm
> -
> -.macro vsetvlstatic16 w, vlen
> -        .if \w == 2 && \vlen == 128
> -                vsetivli        zero, \w, e16, mf4, ta, ma
> -        .elseif \w <= 4 && \vlen == 128
> -                vsetivli        zero, \w, e16, mf2, ta, ma
> -        .elseif \w <= 8 && \vlen == 128
> -                vsetivli        zero, \w, e16, m1, ta, ma
> -        .elseif \w <= 16 && \vlen == 128
> -                vsetivli        zero, \w, e16, m2, ta, ma
> -        .elseif \w <= 32 && \vlen == 128
> -                li              t0, \w
> -                vsetvli         zero, t0, e16, m4, ta, ma
> -        .elseif \w <= 4 && \vlen == 256
> -                vsetivli        zero, \w, e16, mf4, ta, ma
> -        .elseif \w <= 8 && \vlen == 256
> -                vsetivli        zero, \w, e16, mf2, ta, ma
> -        .elseif \w <= 16 && \vlen == 256
> -                vsetivli        zero, \w, e16, m1, ta, ma
> -        .elseif \w <= 32 && \vlen == 256
> -                li              t0, \w
> -                vsetvli         zero, t0, e16, m2, ta, ma
> -        .elseif \w <= 64 && \vlen == 256
> -                li              t0, \w
> -                vsetvli         zero, t0, e16, m4, ta, ma
> -        .else
> -                li              t0, \w
> -                vsetvli         zero, t0, e16, m8, ta, ma
> -        .endif
> -.endm
> -
> -.macro vsetvlstatic32 w, vlen
> -        .if \w == 2
> -                vsetivli        zero, \w, e32, mf2, ta, ma
> -        .elseif \w <= 4 && \vlen == 128
> -                vsetivli        zero, \w, e32, m1, ta, ma
> -        .elseif \w <= 8 && \vlen == 128
> -                vsetivli        zero, \w, e32, m2, ta, ma
> -        .elseif \w <= 16 && \vlen == 128
> -                vsetivli        zero, \w, e32, m4, ta, ma
> -        .elseif \w <= 4 && \vlen == 256
> -                vsetivli        zero, \w, e32, mf2, ta, ma
> -        .elseif \w <= 8 && \vlen == 256
> -                vsetivli        zero, \w, e32, m1, ta, ma
> -        .elseif \w <= 16 && \vlen == 256
> -                vsetivli        zero, \w, e32, m2, ta, ma
> -        .elseif \w <= 32 && \vlen == 256
> -                li              t0, \w
> -                vsetvli         zero, t0, e32, m4, ta, ma
> -        .else
> -                li              t0, \w
> -                vsetvli         zero, t0, e32, m8, ta, ma
> -        .endif
> -.endm
> +#include "libavcodec/riscv/h26x/asm.S"
> 
>  .macro avg w, vlen, id
>  \id\w\vlen:
> @@ -157,35 +73,12 @@
>          ret
>  .endm
> 
> -
> -.macro AVG_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
> -        .4byte \id\()8\vlen\()f - jmp_table_\id\vlen
> -        .4byte \id\()16\vlen\()f - jmp_table_\id\vlen
> -        .4byte \id\()32\vlen\()f - jmp_table_\id\vlen
> -        .4byte \id\()64\vlen\()f - jmp_table_\id\vlen
> -        .4byte \id\()128\vlen\()f - jmp_table_\id\vlen
> -endconst
> -.endm
> -
> -.macro AVG_J vlen, id
> -        clz               t1, a4
> -        neg               t1, t1
> -        lla               t5, jmp_table_\id\vlen
> -        sh2add            t1, t1, t5
> -        lw                t1, ((__riscv_xlen-2)<<2)(t1)
> -        add               t1, t1, t5
> -        jr                t1
> -.endm
> -
>  .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 +158,7 @@ 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 +168,7 @@ 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
> @@ -407,3 +300,6 @@ func_dmvr \vlen, dmvr_h
>  func_dmvr \vlen, dmvr_v
>  func_dmvr \vlen, dmvr_hv
>  .endr
> +
> +func_put_pixels 256, 128, vvc
> +func_put_pixels 128, 128, vvc
> diff --git a/libavcodec/riscv/vvc/vvcdsp_init.c
> b/libavcodec/riscv/vvc/vvcdsp_init.c index aa5ab83a45..2fe93029aa 100644
> --- a/libavcodec/riscv/vvc/vvcdsp_init.c
> +++ b/libavcodec/riscv/vvc/vvcdsp_init.c
> @@ -24,6 +24,8 @@
>  #include "libavutil/cpu.h"
>  #include "libavutil/riscv/cpu.h"
>  #include "libavcodec/vvc/dsp.h"
> +#include "libavcodec/vvc/dec.h"
> +#include "libavcodec/riscv/h26x/h2656dsp.h"
> 
>  #define bf(fn, bd,  opt) fn##_##bd##_##opt
> 
> @@ -57,6 +59,23 @@ 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
> @@ -75,6 +94,8 @@ 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;
> @@ -87,6 +108,8 @@ 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;


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





More information about the ffmpeg-devel mailing list