[FFmpeg-devel] [PATCH 1/2] lavc/aacpsdsp: rework RISC-V add_squares

[Rémi Denis-Courmont]

Le lauantaina 15. heinäkuuta 2023, 23.57.04 EEST Rémi Denis-Courmont a écrit :
> Segmented loads are kinda slow, so this advantageously uses a
> unit-strided load and narrowing shifts instead.
> 
> Before:
> ps_add_squares_rvv_f32: 30352.2
> 
> After:
> ps_add_squares_rvv_f32: 11973.7
> ---
>  libavcodec/riscv/aacpsdsp_rvv.S | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/libavcodec/riscv/aacpsdsp_rvv.S
> b/libavcodec/riscv/aacpsdsp_rvv.S index 80bd19f6ad..b7ea314fc3 100644
> --- a/libavcodec/riscv/aacpsdsp_rvv.S
> +++ b/libavcodec/riscv/aacpsdsp_rvv.S
> @@ -1,5 +1,5 @@
>  /*
> - * Copyright © 2022 Rémi Denis-Courmont.
> + * Copyright © 2022-2023 Rémi Denis-Courmont.
>   *
>   * This file is part of FFmpeg.
>   *
> @@ -21,13 +21,16 @@
>  #include "libavutil/riscv/asm.S"
> 
>  func ff_ps_add_squares_rvv, zve32f
> +        li          t1, 32
>  1:
>          vsetvli     t0, a2, e32, m1, ta, ma
> -        vlseg2e32.v v24, (a1)
> -        sub         a2, a2, t0
> +        vle64.v     v8, (a1)

Requires zve64x

> +        vnsrl.wx    v24, v8, zero
>          vle32.v     v16, (a0)
> -        sh3add      a1, t0, a1
> +        sub         a2, a2, t0
> +        vnsrl.wx    v25, v8, t1
>          vfmacc.vv   v16, v24, v24
> +        sh3add      a1, t0, a1
>          vfmacc.vv   v16, v25, v25
>          vse32.v     v16, (a0)
>          sh2add      a0, t0, a0

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