[FFmpeg-devel] [PATCH v2] lavc/vvc_mc: R-V V avg w_avg

Rémi Denis-Courmont remi at remlab.net
Sat Jun 1 22:54:10 EEST 2024


Le lauantaina 1. kesäkuuta 2024, 21.01.16 EEST uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> 
>                                                       C908   X60
> avg_8_2x2_c                                        :    1.0    1.0
> avg_8_2x2_rvv_i32                                  :    1.0    1.0

I think we can drop the 2x2 transforms. In all likelihood, scalar code will 
end up faster than vector code on future hardware, especially out-of-order 
pipelines.

> avg_8_2x4_c                                        :    1.7    2.0
> avg_8_2x4_rvv_i32                                  :    1.2    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    7.5
> avg_8_2x16_rvv_i32                                 :    3.2    3.0
> avg_8_2x32_c                                       :   14.2   15.0
> avg_8_2x32_rvv_i32                                 :    5.7    5.0
> avg_8_2x64_c                                       :   46.7   44.2
> avg_8_2x64_rvv_i32                                 :   39.2   36.0
> avg_8_2x128_c                                      :   99.7   80.0
> avg_8_2x128_rvv_i32                                :   86.2   65.5
> avg_8_4x2_c                                        :    2.0    2.0
> avg_8_4x2_rvv_i32                                  :    1.0    1.0
> avg_8_4x4_c                                        :    3.5    3.7
> avg_8_4x4_rvv_i32                                  :    1.5    1.2
> avg_8_4x8_c                                        :    6.5    7.0
> avg_8_4x8_rvv_i32                                  :    2.0    1.7
> avg_8_4x16_c                                       :   13.5   14.0
> avg_8_4x16_rvv_i32                                 :    3.2    2.7
> avg_8_4x32_c                                       :   26.2   27.5
> avg_8_4x32_rvv_i32                                 :    5.7    5.0
> avg_8_4x64_c                                       :   75.0   65.7
> avg_8_4x64_rvv_i32                                 :   44.0   32.0
> avg_8_4x128_c                                      :  165.2  118.5
> avg_8_4x128_rvv_i32                                :   81.5   71.0
> avg_8_8x2_c                                        :    3.2    3.5
> avg_8_8x2_rvv_i32                                  :    1.2    1.0
> avg_8_8x4_c                                        :    6.5    6.5
> avg_8_8x4_rvv_i32                                  :    1.5    1.5
> avg_8_8x8_c                                        :   12.5   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                                       :   50.0   52.5
> avg_8_8x32_rvv_i32                                 :    6.7    5.2
> avg_8_8x64_c                                       :  120.7  119.0
> avg_8_8x64_rvv_i32                                 :   43.2   33.5
> avg_8_8x128_c                                      :  247.5  217.7
> avg_8_8x128_rvv_i32                                :  100.5   74.7
> avg_8_16x2_c                                       :    6.2    6.5
> avg_8_16x2_rvv_i32                                 :    1.2    1.0
> avg_8_16x4_c                                       :   12.2   13.0
> avg_8_16x4_rvv_i32                                 :    2.0    1.2
> avg_8_16x8_c                                       :   24.5   25.7
> avg_8_16x8_rvv_i32                                 :    3.2    2.0
> avg_8_16x16_c                                      :   48.7   51.2
> avg_8_16x16_rvv_i32                                :    5.7    3.2
> avg_8_16x32_c                                      :   97.5  102.7
> avg_8_16x32_rvv_i32                                :   10.7    6.0
> avg_8_16x64_c                                      :  213.0  215.0
> avg_8_16x64_rvv_i32                                :   51.5   33.5
> avg_8_16x128_c                                     :  408.5  417.0
> avg_8_16x128_rvv_i32                               :  102.0   71.5
> avg_8_32x2_c                                       :   12.2   13.0
> 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    3.0
> avg_8_32x16_c                                      :   96.5  101.5
> avg_8_32x16_rvv_i32                                :   10.5    5.0
> avg_8_32x32_c                                      :  210.2  202.5
> avg_8_32x32_rvv_i32                                :   20.2    9.7
> avg_8_32x64_c                                      :  431.7  417.2
> avg_8_32x64_rvv_i32                                :   68.0   46.0
> avg_8_32x128_c                                     :  822.2  819.0
> avg_8_32x128_rvv_i32                               :  152.2   69.0
> avg_8_64x2_c                                       :   24.0   25.2
> avg_8_64x2_rvv_i32                                 :    3.0    1.7
> avg_8_64x4_c                                       :   48.2   51.0
> avg_8_64x4_rvv_i32                                 :    5.5    2.7
> avg_8_64x8_c                                       :   96.7  101.5
> avg_8_64x8_rvv_i32                                 :   10.0    5.0
> avg_8_64x16_c                                      :  193.5  203.0
> avg_8_64x16_rvv_i32                                :   19.2    9.2
> avg_8_64x32_c                                      :  404.2  405.7
> avg_8_64x32_rvv_i32                                :   37.7   18.0
> avg_8_64x64_c                                      :  846.2  841.7
> avg_8_64x64_rvv_i32                                :  136.5   35.2
> avg_8_64x128_c                                     : 1659.5 1662.2
> avg_8_64x128_rvv_i32                               :  236.7  177.7
> avg_8_128x2_c                                      :   48.7   51.0
> 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                                      :  226.0  201.5
> avg_8_128x8_rvv_i32                                :   18.7    8.7
> avg_8_128x16_c                                     :  402.7  402.7
> avg_8_128x16_rvv_i32                               :   37.0   17.2
> avg_8_128x32_c                                     :  791.2  805.0
> avg_8_128x32_rvv_i32                               :   73.5   33.5
> avg_8_128x64_c                                     : 1616.7 1645.2
> avg_8_128x64_rvv_i32                               :  223.7   68.2
> avg_8_128x128_c                                    : 3202.0 3235.2
> avg_8_128x128_rvv_i32                              :  390.0  314.2
> w_avg_8_2x2_c                                      :    1.7    1.5
> w_avg_8_2x2_rvv_i32                                :    1.7    1.5
> w_avg_8_2x4_c                                      :    2.7    2.5
> w_avg_8_2x4_rvv_i32                                :    2.7    2.5
> w_avg_8_2x8_c                                      :    5.0    5.0
> w_avg_8_2x8_rvv_i32                                :    4.5    4.0
> w_avg_8_2x16_c                                     :   26.5    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                                     :   58.5   46.5
> w_avg_8_2x64_rvv_i32                               :   49.7   38.7
> w_avg_8_2x128_c                                    :  121.7   85.2
> w_avg_8_2x128_rvv_i32                              :   89.7   81.0
> w_avg_8_4x2_c                                      :    2.5    2.5
> w_avg_8_4x2_rvv_i32                                :    1.7    1.5
> w_avg_8_4x4_c                                      :    4.7    4.7
> 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                               :   15.0   13.5
> w_avg_8_4x64_c                                     :   95.2   80.2
> w_avg_8_4x64_rvv_i32                               :   47.7   38.0
> w_avg_8_4x128_c                                    :  197.7  164.7
> w_avg_8_4x128_rvv_i32                              :  101.7   81.5
> w_avg_8_8x2_c                                      :    4.5    4.5
> w_avg_8_8x2_rvv_i32                                :    2.0    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                                      :   33.5   17.0
> w_avg_8_8x8_rvv_i32                                :    4.7    4.0
> w_avg_8_8x16_c                                     :   34.0   34.0
> w_avg_8_8x16_rvv_i32                               :    8.5    7.2
> w_avg_8_8x32_c                                     :   85.5   67.7
> w_avg_8_8x32_rvv_i32                               :   16.2   13.5
> w_avg_8_8x64_c                                     :  162.5  148.2
> w_avg_8_8x64_rvv_i32                               :   50.0   36.5
> w_avg_8_8x128_c                                    :  380.2  301.5
> w_avg_8_8x128_rvv_i32                              :   87.2   79.5
> w_avg_8_16x2_c                                     :    8.5    8.7
> w_avg_8_16x2_rvv_i32                               :    2.2    1.7
> w_avg_8_16x4_c                                     :   16.7   17.0
> w_avg_8_16x4_rvv_i32                               :    3.7    2.5
> w_avg_8_16x8_c                                     :   33.2   33.7
> w_avg_8_16x8_rvv_i32                               :    6.5    4.2
> w_avg_8_16x16_c                                    :   66.2   66.5
> w_avg_8_16x16_rvv_i32                              :   12.0    7.5
> w_avg_8_16x32_c                                    :  133.2  134.0
> w_avg_8_16x32_rvv_i32                              :   23.0   14.2
> w_avg_8_16x64_c                                    :  296.0  276.7
> w_avg_8_16x64_rvv_i32                              :   66.7   38.2
> w_avg_8_16x128_c                                   :  625.2  539.7
> w_avg_8_16x128_rvv_i32                             :  135.5   79.2
> 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    3.5
> w_avg_8_32x8_c                                     :   65.7   66.2
> w_avg_8_32x8_rvv_i32                               :   11.2    5.7
> w_avg_8_32x16_c                                    :  132.0  132.0
> w_avg_8_32x16_rvv_i32                              :   21.5   10.7
> w_avg_8_32x32_c                                    :  261.7  272.2
> w_avg_8_32x32_rvv_i32                              :   42.2   20.5
> w_avg_8_32x64_c                                    :  528.2  562.7
> w_avg_8_32x64_rvv_i32                              :   83.5   59.2
> w_avg_8_32x128_c                                   : 1135.5 1070.0
> w_avg_8_32x128_rvv_i32                             :  208.7   96.5
> w_avg_8_64x2_c                                     :   33.0   33.0
> w_avg_8_64x2_rvv_i32                               :    6.0    3.0
> w_avg_8_64x4_c                                     :   65.5   67.0
> w_avg_8_64x4_rvv_i32                               :   11.0    5.2
> w_avg_8_64x8_c                                     :  150.0  134.7
> w_avg_8_64x8_rvv_i32                               :   21.5   10.0
> w_avg_8_64x16_c                                    :  265.2  273.7
> w_avg_8_64x16_rvv_i32                              :   42.2   19.0
> w_avg_8_64x32_c                                    :  629.7  541.7
> w_avg_8_64x32_rvv_i32                              :   83.7   37.7
> w_avg_8_64x64_c                                    : 1259.0 1237.7
> w_avg_8_64x64_rvv_i32                              :  190.7   76.0
> w_avg_8_64x128_c                                   : 2967.0 2209.5
> w_avg_8_64x128_rvv_i32                             :  437.0  190.5
> w_avg_8_128x2_c                                    :   65.7   66.0
> w_avg_8_128x2_rvv_i32                              :   11.2    5.5
> w_avg_8_128x4_c                                    :  131.7  134.7
> w_avg_8_128x4_rvv_i32                              :   21.5   10.0
> w_avg_8_128x8_c                                    :  270.2  264.2
> w_avg_8_128x8_rvv_i32                              :   42.2   19.2
> w_avg_8_128x16_c                                   :  580.0  554.2
> w_avg_8_128x16_rvv_i32                             :   83.5   37.5
> w_avg_8_128x32_c                                   : 1141.0 1206.2
> w_avg_8_128x32_rvv_i32                             :  166.2   74.2
> w_avg_8_128x64_c                                   : 2295.5 2403.5
> w_avg_8_128x64_rvv_i32                             :  408.2  159.0
> w_avg_8_128x128_c                                  : 5367.5 4915.2
> w_avg_8_128x128_rvv_i32                            :  741.2  331.5
> 
> test
> 
> u makefile

Squash left overs?

> ---
>  libavcodec/riscv/vvc/Makefile      |   2 +
>  libavcodec/riscv/vvc/vvc_mc_rvv.S  | 295 +++++++++++++++++++++++++++++
>  libavcodec/riscv/vvc/vvcdsp_init.c |  71 +++++++
>  libavcodec/vvc/dsp.c               |   4 +-
>  libavcodec/vvc/dsp.h               |   1 +
>  5 files changed, 372 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..f2128fa776
> --- /dev/null
> +++ b/libavcodec/riscv/vvc/vvc_mc_rvv.S
> @@ -0,0 +1,295 @@
> +/*
> + * 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
> +        .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
> +    .8byte \id\()2\vlen\()f
> +    .8byte \id\()4\vlen\()f
> +    .8byte \id\()8\vlen\()f
> +    .8byte \id\()16\vlen\()f
> +    .8byte \id\()32\vlen\()f
> +    .8byte \id\()64\vlen\()f
> +    .8byte \id\()128\vlen\()f

AFAIU, this will generate relocations. I wonder if the linker smart enough to 
put that into .data.relro rather than whine that it can't live it in .rodata?

In assembler, we can dodge the problem entirely by storing relative offsets 
rather than addresses. You can also stick to 4- or even 2-byte values then. 

> +endconst
> +.endm
> +
> +.macro AVG_J vlen, id
> +        clz               a4, a4
> +        li                t0, __riscv_xlen-2
> +        sub               a4, t0, a4
> +        lla               t0, jmp_table_\id\vlen
> +        sh3add            t0, a4, t0
> +        ld                t0, 0(t0)

LLA is an alias for AUIPC; ADD. You can avoid that ADD by folding the low bits 
into LD. See how ff_h263_loop_filter_strength is addressed in h263dsp_rvv.S.

> +        jr                t0
> +.endm
> +
> +.macro func_avg vlen
> +func ff_vvc_avg_8_rvv_\vlen\(), zve32x
> +        AVG_JMP_TABLE     1, \vlen
> +        csrw              vxrm, zero
> +        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
> +        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
> +        .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
> +        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
> +        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)) {
> +        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