Re: [FFmpeg-devel] [PATCH 2/6] lavc/ac3dsp: RISC-V V float_to_fixed24

["Rémi Denis-Courmont" <remi@remlab.net>]

Le torstaina 15. kesäkuuta 2023, 13.36.41 EEST Peiting Shen a écrit :
> From: Shen Peiting <shenpeiting@eswincomputing.com>
> 
> Vector instructions replaces scalar options of float convert to fixed
> 
> Benchmarks on Spike(cycles):
> len=16
> float_to_fixed24_c: 315
> float_to_fixed24_rvv: 27
> len=160
> float_to_fixed24_c: 2871
> float_to_fixed24_rvv: 67
> 
> Co-Authored by: Yang Xiaojun <yangxiaojun@eswincomputing.com>
> Co-Authored by: Huang Xing <huangxing1@eswincomputing.com>
> Co-Authored by: Zeng Fanchen <zengfanchen@eswincomputing.com>
> Signed-off-by: Shen Peiting <shenpeiting@eswincomputing.com>
> ---
>  libavcodec/riscv/ac3dsp_init.c |  5 ++++-
>  libavcodec/riscv/ac3dsp_rvv.S  | 19 +++++++++++++++++++
>  2 files changed, 23 insertions(+), 1 deletion(-)
> 
> diff --git a/libavcodec/riscv/ac3dsp_init.c b/libavcodec/riscv/ac3dsp_init.c
> index bb67d86998..a4e75a7541 100644
> --- a/libavcodec/riscv/ac3dsp_init.c
> +++ b/libavcodec/riscv/ac3dsp_init.c
> @@ -25,13 +25,16 @@
>  #include "config.h"
> 
>  void ff_ac3_exponent_min_rvv(uint8_t *exp, int num_reuse_blocks, int
> nb_coefs); +void ff_float_to_fixed24_rvv(int32_t *dst, const float *src,
> unsigned int len);
> 
>  av_cold void ff_ac3dsp_init_riscv(AC3DSPContext *c)
>  {
>      int flags = av_get_cpu_flags();
>  #if HAVE_RVV
> -    if (flags & AV_CPU_FLAG_RVV_I32)
> +    if (flags & AV_CPU_FLAG_RVV_I32) {
>          c->ac3_exponent_min = ff_ac3_exponent_min_rvv;
> +        c->float_to_fixed24 = ff_float_to_fixed24_rvv;
> +    }
>  #endif
>  }
> 
> diff --git a/libavcodec/riscv/ac3dsp_rvv.S b/libavcodec/riscv/ac3dsp_rvv.S
> index 879123f4a7..d98e72c12c 100644
> --- a/libavcodec/riscv/ac3dsp_rvv.S
> +++ b/libavcodec/riscv/ac3dsp_rvv.S
> @@ -44,3 +44,22 @@ func ff_ac3_exponent_min_rvv, zve32x
>  3:
>      ret
>  endfunc
> +
> +
> +func ff_float_to_fixed24_rvv, zve32x
> +    addi            t1, x0, 1

That's `li t1, 1` please.

> +    slli            t1, t1, 24
> +    fcvt.s.w        f1, t1

Please use ABI names for FPRs, e.g. `ft0`. Nobody wants to have to remember 
which ones are callee-saved and which ones aren't.

> +1:
> +    vsetvli         t0, a2, e32, m8
> +    vle32.v         v0, (a1)
> +    vfmul.vf        v0, v0, f1
> +    vfcvt.x.f.v     v16, v0
> +    vse32.v         v16, (a0)
> +    sub             a2, a2, t0
> +    slli            t0, t0, 2
> +    add             a1, a1, t0
> +    add             a0, a0, t0

Use sh2add to save one in three instruction here.

And please interleave scalar and vector instructions so in-order CPU can 
potentially multi-issue.

> +    bgtz            a2, 1b
> +    ret
> +endfunc

-- 
Реми Дёни-Курмон
http://www.remlab.net/



_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".