[FFmpeg-devel] [PATCH v3] lavc/vvc_mc: R-V V avg w_avg

Rémi Denis-Courmont remi at remlab.net
Tue Jun 11 20:06:11 EEST 2024


Le tiistaina 11. kesäkuuta 2024, 19.38.15 EEST uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> 
>                                                       C908   X60
> avg_8_2x2_c                                        :    1.2    1.0
> avg_8_2x2_rvv_i32                                  :    1.0    1.0
> avg_8_2x4_c                                        :    2.0    2.0
> avg_8_2x4_rvv_i32                                  :    1.5    1.2
> avg_8_2x8_c                                        :    3.7    4.0
> avg_8_2x8_rvv_i32                                  :    2.0    2.0
> avg_8_2x16_c                                       :    7.2    7.7
> avg_8_2x16_rvv_i32                                 :    3.2    3.0
> avg_8_2x32_c                                       :   14.5   15.2
> avg_8_2x32_rvv_i32                                 :    5.7    5.0
> avg_8_2x64_c                                       :   50.0   45.2
> avg_8_2x64_rvv_i32                                 :   41.5   32.5
> avg_8_2x128_c                                      :  101.5   84.2
> avg_8_2x128_rvv_i32                                :   89.5   73.2
> avg_8_4x2_c                                        :    2.0    2.0
> avg_8_4x2_rvv_i32                                  :    1.0    1.0
> avg_8_4x4_c                                        :    3.5    3.5
> avg_8_4x4_rvv_i32                                  :    1.5    1.2
> avg_8_4x8_c                                        :    6.7    7.0
> avg_8_4x8_rvv_i32                                  :    2.0    1.7
> avg_8_4x16_c                                       :   13.2   14.0
> avg_8_4x16_rvv_i32                                 :    3.2    3.0
> avg_8_4x32_c                                       :   26.2   27.7
> avg_8_4x32_rvv_i32                                 :    5.7    5.0
> avg_8_4x64_c                                       :   75.0   66.0
> avg_8_4x64_rvv_i32                                 :   40.2   33.0
> avg_8_4x128_c                                      :  144.5  128.0
> avg_8_4x128_rvv_i32                                :   89.5   78.7
> avg_8_8x2_c                                        :    3.2    3.5
> avg_8_8x2_rvv_i32                                  :    1.2    1.0
> avg_8_8x4_c                                        :    6.5    6.7
> avg_8_8x4_rvv_i32                                  :    1.5    1.5
> avg_8_8x8_c                                        :   12.7   13.2
> avg_8_8x8_rvv_i32                                  :    2.2    1.7
> avg_8_8x16_c                                       :   25.2   26.5
> avg_8_8x16_rvv_i32                                 :    3.7    2.7
> avg_8_8x32_c                                       :   50.2   52.7
> avg_8_8x32_rvv_i32                                 :    6.5    5.0
> avg_8_8x64_c                                       :  120.2  117.7
> avg_8_8x64_rvv_i32                                 :   45.2   39.2
> avg_8_8x128_c                                      :  223.0  233.5
> avg_8_8x128_rvv_i32                                :   80.0   73.2
> avg_8_16x2_c                                       :    6.2    6.5
> avg_8_16x2_rvv_i32                                 :    1.5    1.0
> avg_8_16x4_c                                       :   12.5   12.7
> avg_8_16x4_rvv_i32                                 :    2.0    1.2
> avg_8_16x8_c                                       :   24.7   26.0
> avg_8_16x8_rvv_i32                                 :    3.2    2.0
> avg_8_16x16_c                                      :   49.0   51.2
> avg_8_16x16_rvv_i32                                :    5.7    3.2
> avg_8_16x32_c                                      :   97.7  102.5
> avg_8_16x32_rvv_i32                                :   10.7    5.7
> avg_8_16x64_c                                      :  220.5  214.2
> avg_8_16x64_rvv_i32                                :   48.2   39.5
> avg_8_16x128_c                                     :  436.2  428.0
> avg_8_16x128_rvv_i32                               :   97.2   77.0
> avg_8_32x2_c                                       :   12.2   12.7
> avg_8_32x2_rvv_i32                                 :    2.0    1.2
> avg_8_32x4_c                                       :   24.5   25.5
> avg_8_32x4_rvv_i32                                 :    3.2    1.7
> avg_8_32x8_c                                       :   48.5   50.7
> avg_8_32x8_rvv_i32                                 :    5.7    2.7
> avg_8_32x16_c                                      :   96.5  101.2
> avg_8_32x16_rvv_i32                                :   10.2    5.0
> avg_8_32x32_c                                      :  192.5  202.2
> avg_8_32x32_rvv_i32                                :   20.0    9.5
> avg_8_32x64_c                                      :  405.7  404.5
> avg_8_32x64_rvv_i32                                :   72.5   40.2
> avg_8_32x128_c                                     :  821.0  832.2
> avg_8_32x128_rvv_i32                               :  136.2   75.7
> avg_8_64x2_c                                       :   24.0   25.2
> avg_8_64x2_rvv_i32                                 :    3.2    1.7
> avg_8_64x4_c                                       :   48.5   51.0
> avg_8_64x4_rvv_i32                                 :    5.5    2.7
> avg_8_64x8_c                                       :   97.0  101.5
> avg_8_64x8_rvv_i32                                 :   10.2    5.0
> avg_8_64x16_c                                      :  193.5  202.7
> avg_8_64x16_rvv_i32                                :   19.2    9.2
> avg_8_64x32_c                                      :  404.2  405.7
> avg_8_64x32_rvv_i32                                :   38.0   17.7
> avg_8_64x64_c                                      :  834.0  840.7
> avg_8_64x64_rvv_i32                                :   75.0   36.2
> avg_8_64x128_c                                     : 1667.2 1685.7
> avg_8_64x128_rvv_i32                               :  336.0  181.5
> avg_8_128x2_c                                      :   49.0   50.7
> avg_8_128x2_rvv_i32                                :    5.2    2.7
> avg_8_128x4_c                                      :   96.7  101.0
> avg_8_128x4_rvv_i32                                :    9.7    4.7
> avg_8_128x8_c                                      :  193.0  201.7
> avg_8_128x8_rvv_i32                                :   19.0    8.5
> avg_8_128x16_c                                     :  386.2  402.7
> avg_8_128x16_rvv_i32                               :   37.2   79.0
> avg_8_128x32_c                                     :  789.2  805.0
> avg_8_128x32_rvv_i32                               :   73.5   32.7
> avg_8_128x64_c                                     : 1620.5 1651.2
> avg_8_128x64_rvv_i32                               :  185.2   69.7
> avg_8_128x128_c                                    : 3203.0 3236.7
> avg_8_128x128_rvv_i32                              :  457.2  280.5
> w_avg_8_2x2_c                                      :    1.5    1.5
> w_avg_8_2x2_rvv_i32                                :    1.7    1.5
> w_avg_8_2x4_c                                      :    2.7    2.7
> w_avg_8_2x4_rvv_i32                                :    2.7    2.5
> w_avg_8_2x8_c                                      :    5.0    4.7
> w_avg_8_2x8_rvv_i32                                :    4.5    4.0
> w_avg_8_2x16_c                                     :    9.7    9.5
> w_avg_8_2x16_rvv_i32                               :    8.0    7.0
> w_avg_8_2x32_c                                     :   18.7   18.5
> w_avg_8_2x32_rvv_i32                               :   15.0   13.2
> w_avg_8_2x64_c                                     :   57.7   49.0
> w_avg_8_2x64_rvv_i32                               :   42.7   35.2
> w_avg_8_2x128_c                                    :  127.5   94.5
> w_avg_8_2x128_rvv_i32                              :   99.2   78.2
> w_avg_8_4x2_c                                      :    2.5    2.5
> w_avg_8_4x2_rvv_i32                                :    2.0    1.7
> w_avg_8_4x4_c                                      :    4.7    4.5
> w_avg_8_4x4_rvv_i32                                :    2.7    2.2
> w_avg_8_4x8_c                                      :    9.0    9.0
> w_avg_8_4x8_rvv_i32                                :    4.5    4.0
> w_avg_8_4x16_c                                     :   17.7   17.7
> w_avg_8_4x16_rvv_i32                               :    8.0    7.0
> w_avg_8_4x32_c                                     :   35.0   35.0
> w_avg_8_4x32_rvv_i32                               :   32.7   13.2
> w_avg_8_4x64_c                                     :  117.5   79.5
> w_avg_8_4x64_rvv_i32                               :   47.2   39.0
> w_avg_8_4x128_c                                    :  235.7  159.0
> w_avg_8_4x128_rvv_i32                              :  101.0   80.0
> w_avg_8_8x2_c                                      :    4.5    4.5
> w_avg_8_8x2_rvv_i32                                :    1.7    1.7
> w_avg_8_8x4_c                                      :    8.7    8.7
> w_avg_8_8x4_rvv_i32                                :    2.7    2.5
> w_avg_8_8x8_c                                      :   17.2   17.0
> w_avg_8_8x8_rvv_i32                                :    4.7    4.0
> w_avg_8_8x16_c                                     :   34.0   34.2
> w_avg_8_8x16_rvv_i32                               :    8.5    7.0
> w_avg_8_8x32_c                                     :   67.5   67.7
> w_avg_8_8x32_rvv_i32                               :   16.0   13.2
> w_avg_8_8x64_c                                     :  184.0  147.7
> w_avg_8_8x64_rvv_i32                               :   53.7   35.0
> w_avg_8_8x128_c                                    :  350.0  320.5
> w_avg_8_8x128_rvv_i32                              :   98.5   74.7
> w_avg_8_16x2_c                                     :    8.7    8.5
> w_avg_8_16x2_rvv_i32                               :    2.5    1.7
> w_avg_8_16x4_c                                     :   17.0   17.0
> w_avg_8_16x4_rvv_i32                               :    3.7    2.5
> w_avg_8_16x8_c                                     :   49.7   33.5
> w_avg_8_16x8_rvv_i32                               :    6.5    4.2
> w_avg_8_16x16_c                                    :   66.5   66.5
> w_avg_8_16x16_rvv_i32                              :   12.2    7.5
> w_avg_8_16x32_c                                    :  132.2  134.0
> w_avg_8_16x32_rvv_i32                              :   23.2   14.2
> w_avg_8_16x64_c                                    :  298.2  283.5
> w_avg_8_16x64_rvv_i32                              :   65.7   40.2
> w_avg_8_16x128_c                                   :  755.5  593.2
> w_avg_8_16x128_rvv_i32                             :  132.0   76.0
> w_avg_8_32x2_c                                     :   16.7   16.7
> w_avg_8_32x2_rvv_i32                               :    3.5    2.0
> w_avg_8_32x4_c                                     :   33.2   33.2
> w_avg_8_32x4_rvv_i32                               :    6.0    3.2
> w_avg_8_32x8_c                                     :   65.7   66.0
> w_avg_8_32x8_rvv_i32                               :   11.2    6.5
> w_avg_8_32x16_c                                    :  148.2  132.0
> w_avg_8_32x16_rvv_i32                              :   21.5   10.7
> w_avg_8_32x32_c                                    :  266.2  267.0
> w_avg_8_32x32_rvv_i32                              :   60.7   20.7
> w_avg_8_32x64_c                                    :  683.5  559.7
> w_avg_8_32x64_rvv_i32                              :   83.7   65.0
> w_avg_8_32x128_c                                   : 1169.5 1140.2
> w_avg_8_32x128_rvv_i32                             :  191.2   96.7
> w_avg_8_64x2_c                                     :   33.0   33.2
> w_avg_8_64x2_rvv_i32                               :    6.0    3.2
> w_avg_8_64x4_c                                     :   65.5   65.7
> w_avg_8_64x4_rvv_i32                               :   11.2    5.2
> w_avg_8_64x8_c                                     :  149.7  132.0
> w_avg_8_64x8_rvv_i32                               :   21.7    9.7
> w_avg_8_64x16_c                                    :  279.2  262.7
> w_avg_8_64x16_rvv_i32                              :   42.2   18.5
> w_avg_8_64x32_c                                    :  538.7  542.2
> w_avg_8_64x32_rvv_i32                              :   83.7   36.5
> w_avg_8_64x64_c                                    : 1200.2 1074.2
> w_avg_8_64x64_rvv_i32                              :  204.5   73.7
> w_avg_8_64x128_c                                   : 2375.7 2482.0
> w_avg_8_64x128_rvv_i32                             :  390.5  205.2
> w_avg_8_128x2_c                                    :   66.2   66.5
> w_avg_8_128x2_rvv_i32                              :   11.0    5.2
> w_avg_8_128x4_c                                    :  133.2  133.2
> w_avg_8_128x4_rvv_i32                              :   21.7   10.0
> w_avg_8_128x8_c                                    :  303.5  268.5
> w_avg_8_128x8_rvv_i32                              :   42.2   18.5
> w_avg_8_128x16_c                                   :  544.0  545.5
> w_avg_8_128x16_rvv_i32                             :   83.5   36.5
> w_avg_8_128x32_c                                   : 1128.0 1090.7
> w_avg_8_128x32_rvv_i32                             :  166.5   72.2
> w_avg_8_128x64_c                                   : 2275.7 2167.5
> w_avg_8_128x64_rvv_i32                             :  391.7  146.7
> w_avg_8_128x128_c                                  : 4851.2 4310.5
> w_avg_8_128x128_rvv_i32                            :  742.0  341.2
> ---
>  libavcodec/riscv/vvc/Makefile      |   2 +
>  libavcodec/riscv/vvc/vvc_mc_rvv.S  | 296 +++++++++++++++++++++++++++++
>  libavcodec/riscv/vvc/vvcdsp_init.c |  71 +++++++
>  libavcodec/vvc/dsp.c               |   4 +-
>  libavcodec/vvc/dsp.h               |   1 +
>  5 files changed, 373 insertions(+), 1 deletion(-)
>  create mode 100644 libavcodec/riscv/vvc/Makefile
>  create mode 100644 libavcodec/riscv/vvc/vvc_mc_rvv.S
>  create mode 100644 libavcodec/riscv/vvc/vvcdsp_init.c
> 
> diff --git a/libavcodec/riscv/vvc/Makefile b/libavcodec/riscv/vvc/Makefile
> new file mode 100644
> index 0000000000..582b051579
> --- /dev/null
> +++ b/libavcodec/riscv/vvc/Makefile
> @@ -0,0 +1,2 @@
> +OBJS-$(CONFIG_VVC_DECODER) += riscv/vvc/vvcdsp_init.o
> +RVV-OBJS-$(CONFIG_VVC_DECODER) += riscv/vvc/vvc_mc_rvv.o
> diff --git a/libavcodec/riscv/vvc/vvc_mc_rvv.S
> b/libavcodec/riscv/vvc/vvc_mc_rvv.S new file mode 100644
> index 0000000000..e6e906f3b2
> --- /dev/null
> +++ b/libavcodec/riscv/vvc/vvc_mc_rvv.S
> @@ -0,0 +1,296 @@
> +/*
> + * 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, is_w

Could SEW be a parameter so that these three macros would be a little bit more 
factored? .ifc / .ifnc might help to match e8/e16/e32.

> +        .if \w <= 2
> +                vsetivli        zero, \w, e8, mf8, ta, ma
> +        .elseif \w <= 4 && \vlen == 128
> +                vsetivli        zero, \w, e8, mf4, ta, ma
> +        .elseif \w <= 4 && \vlen >= 256
> +                vsetivli        zero, \w, e8, mf8, ta, ma
> +        .elseif \w <= 8 && \vlen == 128
> +                vsetivli        zero, \w, e8, mf2, ta, ma
> +        .elseif \w <= 8 && \vlen >= 256
> +                vsetivli        zero, \w, e8, mf4, ta, ma
> +        .elseif \w <= 16 && \vlen == 128
> +                vsetivli        zero, \w, e8, m1, 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 <= (\vlen / 4) || \is_w
> +                li t0, 64
> +                vsetvli         zero, t0, e8, m2, ta, ma
> +        .else
> +                li t0, \w
> +                vsetvli         zero, t0, e8, m4, ta, ma
> +        .endif
> +.endm
> +
> +.macro vsetvlstatic16 w, vlen, is_w
> +        .if \w <= 2
> +                vsetivli        zero, \w, e16, mf4, ta, ma
> +        .elseif \w <= 4 && \vlen == 128
> +                vsetivli        zero, \w, e16, mf2, ta, ma
> +        .elseif \w <= 4 && \vlen >= 256
> +                vsetivli        zero, \w, e16, mf4, ta, ma
> +        .elseif \w <= 8 && \vlen == 128
> +                vsetivli        zero, \w, e16, m1, ta, ma
> +        .elseif \w <= 8 && \vlen >= 256
> +                vsetivli        zero, \w, e16, mf2, ta, ma
> +        .elseif \w <= 16 && \vlen == 128
> +                vsetivli        zero, \w, e16, m2, 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 <= (\vlen / 4) || \is_w
> +                li t0, 64
> +                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 <= 4 && \vlen >= 256
> +                vsetivli        zero, \w, e32, mf2, ta, ma
> +        .elseif \w <= 8 && \vlen == 128
> +                vsetivli        zero, \w, e32, m2, ta, ma
> +        .elseif \w <= 8 && \vlen >= 256
> +                vsetivli        zero, \w, e32, m1, ta, ma
> +        .elseif \w <= 16 && \vlen == 128
> +                vsetivli        zero, \w, e32, m4, 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
> +
> +.macro avg_nx1 w, vlen
> +        vsetvlstatic16    \w, \vlen, 0
> +        vle16.v           v0, (a2)
> +        vle16.v           v8, (a3)
> +        vadd.vv           v8, v8, v0
> +        vmax.vx           v8, v8, zero
> +        vsetvlstatic8     \w, \vlen, 0
> +        vnclipu.wi        v8, v8, 7
> +        vse8.v            v8, (a0)
> +.endm
> +
> +.macro avg w, vlen, id
> +\id\w\vlen:
> +.if \w < 128
> +        vsetvlstatic16    \w, \vlen, 0
> +        addi              t0, a2, 128*2
> +        addi              t1, a3, 128*2
> +        add               t2, a0, a1
> +        vle16.v           v0, (a2)
> +        vle16.v           v8, (a3)
> +        addi              a5, a5, -2
> +        vle16.v           v16, (t0)
> +        vle16.v           v24, (t1)
> +        vadd.vv           v8, v8, v0
> +        vadd.vv           v24, v24, v16
> +        vmax.vx           v8, v8, zero
> +        vmax.vx           v24, v24, zero
> +        vsetvlstatic8     \w, \vlen, 0
> +        addi              a2, a2, 128*4
> +        vnclipu.wi        v8, v8, 7
> +        vnclipu.wi        v24, v24, 7
> +        addi              a3, a3, 128*4
> +        vse8.v            v8, (a0)
> +        vse8.v            v24, (t2)
> +        sh1add            a0, a1, a0
> +.else
> +        avg_nx1           128, \vlen
> +        addi              a5, a5, -1
> +        .if \vlen == 128
> +        addi              a2, a2, 64*2
> +        addi              a3, a3, 64*2
> +        addi              a0, a0, 64
> +        avg_nx1           128, \vlen
> +        addi              a0, a0, -64
> +        addi              a2, a2, 128
> +        addi              a3, a3, 128
> +        .else
> +        addi              a2, a2, 128*2
> +        addi              a3, a3, 128*2
> +        .endif
> +        add               a0, a0, a1
> +.endif
> +        bnez              a5, \id\w\vlen\()b
> +        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               a4, a4
> +        li                t0, __riscv_xlen-2
> +        sub               a4, t0, a4
> +        lla               t5, jmp_table_\id\vlen

In C, it would be invalid pointer arithmetic, but in assembler, you can add 
whatever constant offset you want to this symbol, even if points outside the 
table. So you should be able to eliminate the LI above. It won't make much 
difference though.

> +        sh2add            t0, a4, t5
> +        lw                t0, 0(t0)
> +        add               t0, t0, t5
> +        jr                t0
> +.endm
> +
> +.macro func_avg vlen
> +func ff_vvc_avg_8_rvv_\vlen\(), zve32x
> +        AVG_JMP_TABLE     1, \vlen
> +        csrw              vxrm, zero

Nit: for overall code base consistency, I'd use csrwi here. Reason being that 
for other rounding modes, csrwi is the better option.

> +        AVG_J             \vlen, 1
> +        .irp w,2,4,8,16,32,64,128
> +        avg               \w, \vlen, 1
> +        .endr
> +endfunc
> +.endm
> +
> +.macro w_avg_nx1 w, vlen
> +        vsetvlstatic16    \w, \vlen, 1
> +        vle16.v           v0, (a2)
> +        vle16.v           v8, (a3)
> +        vwmul.vx          v16, v0, a7
> +        vwmacc.vx         v16, t3, v8
> +        vsetvlstatic32    \w, \vlen
> +        vadd.vx           v16, v16, t4

I guess t4 is 32-bit? Kinda sad to switch VTYPE just for this but if so, I 
don't have any better idea :(

> +        vsetvlstatic16    \w, \vlen, 1
> +        vnsrl.wx          v16, v16, t6
> +        vmax.vx           v16, v16, zero
> +        vsetvlstatic8     \w, \vlen, 1
> +        vnclipu.wi        v16, v16, 0
> +        vse8.v            v16, (a0)
> +.endm
> +
> +.macro w_avg w, vlen, id
> +\id\w\vlen:
> +.if \vlen <= 16
> +        vsetvlstatic16    \w, \vlen, 1
> +        addi              t0, a2, 128*2
> +        addi              t1, a3, 128*2
> +        vle16.v           v0, (a2)
> +        vle16.v           v8, (a3)
> +        addi              a5, a5, -2
> +        vle16.v           v20, (t0)
> +        vle16.v           v24, (t1)
> +        vwmul.vx          v16, v0, a7
> +        vwmul.vx          v28, v20, a7
> +        vwmacc.vx         v16, t3, v8
> +        vwmacc.vx         v28, t3, v24
> +        vsetvlstatic32    \w, \vlen
> +        add               t2, a0, a1
> +        vadd.vx           v16, v16, t4
> +        vadd.vx           v28, v28, t4
> +        vsetvlstatic16    \w, \vlen, 1
> +        vnsrl.wx          v16, v16, t6
> +        vnsrl.wx          v28, v28, t6
> +        vmax.vx           v16, v16, zero
> +        vmax.vx           v28, v28, zero
> +        vsetvlstatic8     \w, \vlen, 1
> +        addi              a2, a2, 128*4
> +        vnclipu.wi        v16, v16, 0
> +        vnclipu.wi        v28, v28, 0
> +        vse8.v            v16, (a0)
> +        addi              a3, a3, 128*4
> +        vse8.v            v28, (t2)
> +        sh1add            a0, a1, a0
> +.else
> +        w_avg_nx1         \w, \vlen
> +        addi              a5, a5, -1
> +        .if \w == (\vlen / 2)
> +        addi              a2, a2, (\vlen / 2)
> +        addi              a3, a3, (\vlen / 2)
> +        addi              a0, a0, (\vlen / 4)
> +        w_avg_nx1         \w, \vlen
> +        addi              a2, a2, -(\vlen / 2)
> +        addi              a3, a3, -(\vlen / 2)
> +        addi              a0, a0, -(\vlen / 4)
> +        .elseif \w == 128 && \vlen == 128
> +        .rept 3
> +        addi              a2, a2, (\vlen / 2)
> +        addi              a3, a3, (\vlen / 2)
> +        addi              a0, a0, (\vlen / 4)
> +        w_avg_nx1         \w, \vlen

Is that .rept meaningfully faster than a run-time loop?

> +        .endr
> +        addi              a2, a2, -(\vlen / 2) * 3
> +        addi              a3, a3, -(\vlen / 2) * 3
> +        addi              a0, a0, -(\vlen / 4) * 3
> +        .endif
> +
> +        addi              a2, a2, 128*2
> +        addi              a3, a3, 128*2
> +        add               a0, a0, a1
> +.endif
> +        bnez              a5, \id\w\vlen\()b
> +        ret
> +.endm
> +
> +
> +.macro func_w_avg vlen
> +func ff_vvc_w_avg_8_rvv_\vlen\(), zve32x
> +        AVG_JMP_TABLE     2, \vlen
> +        csrw              vxrm, zero

Nit again

> +        addi              t6, a6, 7
> +        ld                t3, (sp)
> +        ld                t4, 8(sp)
> +        ld                t5, 16(sp)
> +        add               t4, t4, t5
> +        addi              t4, t4, 1       // o0 + o1 + 1

Probably faster to swap the two above, to avoid stalling on LD.

> +        addi              t5, t6, -1      // shift - 1
> +        sll               t4, t4, t5
> +        AVG_J             \vlen, 2
> +        .irp w,2,4,8,16,32,64,128
> +        w_avg             \w, \vlen, 2
> +        .endr
> +endfunc
> +.endm
> +
> +func_avg 128
> +func_avg 256
> +#if (__riscv_xlen == 64)
> +func_w_avg 128
> +func_w_avg 256
> +#endif
> diff --git a/libavcodec/riscv/vvc/vvcdsp_init.c
> b/libavcodec/riscv/vvc/vvcdsp_init.c new file mode 100644
> index 0000000000..85b1ede061
> --- /dev/null
> +++ b/libavcodec/riscv/vvc/vvcdsp_init.c
> @@ -0,0 +1,71 @@
> +/*
> + * 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 "config.h"
> +
> +#include "libavutil/attributes.h"
> +#include "libavutil/cpu.h"
> +#include "libavutil/riscv/cpu.h"
> +#include "libavcodec/vvc/dsp.h"
> +
> +#define bf(fn, bd,  opt) fn##_##bd##_##opt
> +
> +#define AVG_PROTOTYPES(bd, opt)                                            
>                          \ +void bf(ff_vvc_avg, bd, opt)(uint8_t *dst,
> ptrdiff_t dst_stride,                                     \ +    const
> int16_t *src0, const int16_t *src1, int width, int height);                
>                \ +void bf(ff_vvc_w_avg, bd, opt)(uint8_t *dst, ptrdiff_t
> dst_stride,                                   \ +    const int16_t *src0,
> const int16_t *src1, int width, int height,                                
> \ +    int denom, int w0, int w1, int o0, int o1);
> +
> +AVG_PROTOTYPES(8, rvv_128)
> +AVG_PROTOTYPES(8, rvv_256)
> +
> +void ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const int bd)
> +{
> +#if HAVE_RVV
> +    const int flags = av_get_cpu_flags();
> +
> +    if ((flags & AV_CPU_FLAG_RVV_I32) && (flags & AV_CPU_FLAG_RVB_ADDR) &&
> +        ff_rv_vlen_least(256)) {

If you check more than one length, better to get ff_get_rv_vlenb() into a local 
variable.

> +        switch (bd) {
> +            case 8:
> +                c->inter.avg    = ff_vvc_avg_8_rvv_256;
> +# if (__riscv_xlen == 64)
> +                c->inter.w_avg    = ff_vvc_w_avg_8_rvv_256;
> +# endif
> +                break;
> +            default:
> +                break;
> +        }
> +    } else if ((flags & AV_CPU_FLAG_RVV_I32) && (flags &
> AV_CPU_FLAG_RVB_ADDR) &&
> +               ff_rv_vlen_least(128)) {
> +        switch (bd) {
> +            case 8:
> +                c->inter.avg    = ff_vvc_avg_8_rvv_128;
> +# if (__riscv_xlen == 64)
> +                c->inter.w_avg    = ff_vvc_w_avg_8_rvv_128;
> +# endif
> +                break;
> +            default:
> +                break;
> +        }
> +    }
> +#endif
> +}
> diff --git a/libavcodec/vvc/dsp.c b/libavcodec/vvc/dsp.c
> index 41e830a98a..c55a37d255 100644
> --- a/libavcodec/vvc/dsp.c
> +++ b/libavcodec/vvc/dsp.c
> @@ -121,7 +121,9 @@ void ff_vvc_dsp_init(VVCDSPContext *vvcdsp, int
> bit_depth) break;
>      }
> 
> -#if ARCH_X86
> +#if ARCH_RISCV
> +    ff_vvc_dsp_init_riscv(vvcdsp, bit_depth);
> +#elif ARCH_X86
>      ff_vvc_dsp_init_x86(vvcdsp, bit_depth);
>  #endif
>  }
> diff --git a/libavcodec/vvc/dsp.h b/libavcodec/vvc/dsp.h
> index 1f14096c41..e03236dd76 100644
> --- a/libavcodec/vvc/dsp.h
> +++ b/libavcodec/vvc/dsp.h
> @@ -180,6 +180,7 @@ typedef struct VVCDSPContext {
> 
>  void ff_vvc_dsp_init(VVCDSPContext *hpc, int bit_depth);
> 
> +void ff_vvc_dsp_init_riscv(VVCDSPContext *hpc, const int bit_depth);
>  void ff_vvc_dsp_init_x86(VVCDSPContext *hpc, const int bit_depth);
> 
>  #endif /* AVCODEC_VVC_DSP_H */


-- 
雷米‧德尼-库尔蒙
http://www.remlab.net/





More information about the ffmpeg-devel mailing list