| Message ID | CAEa-L+v6NmH1vT3_FG=3ciNHTWB_2+_fPQX+N=cG-fuYFMhoTw@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel,1/6] checkasm/takdsp: add decorrelate_ls test | expand |
| Context | Check | Description |
|---|---|---|
| andriy/configure_x86 | warning | Failed to apply patch |
| yinshiyou/configure_loongarch64 | warning | Failed to apply patch |
Le maanantaina 18. joulukuuta 2023, 17.16.27 EET flow gg a écrit : > C908: > decorrelate_sm_c: 130.0 > decorrelate_sm_rvv_i32: 43.7 + +func ff_decorrelate_sm_rvv, zve32x +1: + vsetvli t0, a2, e32, m8, ta, ma + vle32.v v0, (a0) + sub a2, a2, t0 + vle32.v v8, (a1) + vsra.vi v16, v8, 1 You should load v8 first, since it is used as input before v0. + vsub.vv v0, v0, v16 + vse32.v v0, (a0) + sh2add a0, t0, a0 + vadd.vv v0, v0, v8 You can use VSSRA, and then VADD won't need to depend on the output of VSUB. + vse32.v v0, (a1) + sh2add a1, t0, a1 + bnez a2, 1b + ret +endfunc
Le torstaina 21. joulukuuta 2023, 18.07.55 EET Rémi Denis-Courmont a écrit :
> You can use VSSRA, and then VADD won't need to depend on the output of VSUB.
P.S.: I have NOT checked which approach is actually faster.
func ff_decorrelate_sm_rvv, zve32x
1:
vsetvli t0, a2, e32, m8, ta, ma
vle32.v v8, (a1)
sub a2, a2, t0
vle32.v v0, (a0)
vssra.vi v8, v8, 1
vsub.vv v16, v0, v8
vse32.v v16, (a0)
sh2add a0, t0, a0
vadd.vv v16, v0, v8
vse32.v v16, (a1)
sh2add a1, t0, a1
bnez a2, 1b
ret
endfunc
Is this way? In this situation, or when using vsra, there will be some
tests that fail, and the result value differs by 1. I'm not sure where the
problem..
Rémi Denis-Courmont <remi@remlab.net> 于2023年12月22日周五 00:08写道：
> Le maanantaina 18. joulukuuta 2023, 17.16.27 EET flow gg a écrit :
> > C908:
> > decorrelate_sm_c: 130.0
> > decorrelate_sm_rvv_i32: 43.7
>
> +
> +func ff_decorrelate_sm_rvv, zve32x
> +1:
> + vsetvli t0, a2, e32, m8, ta, ma
> + vle32.v v0, (a0)
> + sub a2, a2, t0
> + vle32.v v8, (a1)
> + vsra.vi v16, v8, 1
>
> You should load v8 first, since it is used as input before v0.
>
> + vsub.vv v0, v0, v16
> + vse32.v v0, (a0)
> + sh2add a0, t0, a0
> + vadd.vv v0, v0, v8
>
> You can use VSSRA, and then VADD won't need to depend on the output of
> VSUB.
>
> + vse32.v v0, (a1)
> + sh2add a1, t0, a1
> + bnez a2, 1b
> + ret
> +endfunc
>
> --
> 雷米‧德尼-库尔蒙
> 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".
>
Le perjantaina 22. joulukuuta 2023, 3.34.39 EET flow gg a écrit : > func ff_decorrelate_sm_rvv, zve32x > 1: > vsetvli t0, a2, e32, m8, ta, ma > vle32.v v8, (a1) > sub a2, a2, t0 > vle32.v v0, (a0) > vssra.vi v8, v8, 1 > vsub.vv v16, v0, v8 > vse32.v v16, (a0) > sh2add a0, t0, a0 > vadd.vv v16, v0, v8 > vse32.v v16, (a1) > sh2add a1, t0, a1 > bnez a2, 1b > ret > endfunc > > Is this way? In this situation, or when using vsra, there will be some > tests that fail, and the result value differs by 1. I'm not sure where the > problem.. No, I meant something like this, but it turns out slightly slower anyway. Saving the data dependency is not worth adding an instruction. func ff_decorrelate_sm_rvv, zve32x csrwi vxrm, 0 1: vsetvli t0, a2, e32, m8, ta, ma vle32.v v8, (a1) sub a2, a2, t0 vle32.v v0, (a0) vsra.vi v16, v8, 1 vssra.vi v8, v8, 1 vsub.vv v16, v0, v16 vadd.vv v8, v0, v8 vse32.v v16, (a0) sh2add a0, t0, a0 vse32.v v8, (a1) sh2add a1, t0, a1 bnez a2, 1b ret endfunc
From 3dc613feaa6c38a7df47a3fc385e2140716e0ae2 Mon Sep 17 00:00:00 2001 From: sunyuechi <sunyuechi@iscas.ac.cn> Date: Mon, 18 Dec 2023 22:53:39 +0800 Subject: [PATCH 6/6] lavc/takdsp: R-V V decorrelate_sm C908: decorrelate_sm_c: 130.0 decorrelate_sm_rvv_i32: 43.7 --- libavcodec/riscv/takdsp_init.c | 2 ++ libavcodec/riscv/takdsp_rvv.S | 17 +++++++++++++++++ 2 files changed, 19 insertions(+) diff --git a/libavcodec/riscv/takdsp_init.c b/libavcodec/riscv/takdsp_init.c index 0b4ec18086..85634d6db6 100644 --- a/libavcodec/riscv/takdsp_init.c +++ b/libavcodec/riscv/takdsp_init.c @@ -27,6 +27,7 @@ void ff_decorrelate_ls_rvv(int32_t *p1, int32_t *p2, int length); void ff_decorrelate_sr_rvv(int32_t *p1, int32_t *p2, int length); +void ff_decorrelate_sm_rvv(int32_t *p1, int32_t *p2, int length); av_cold void ff_takdsp_init_riscv(TAKDSPContext *dsp) { @@ -36,6 +37,7 @@ av_cold void ff_takdsp_init_riscv(TAKDSPContext *dsp) if ((flags & AV_CPU_FLAG_RVV_I32) && (flags & AV_CPU_FLAG_RVB_ADDR)) { dsp->decorrelate_ls = ff_decorrelate_ls_rvv; dsp->decorrelate_sr = ff_decorrelate_sr_rvv; + dsp->decorrelate_sm = ff_decorrelate_sm_rvv; } #endif } diff --git a/libavcodec/riscv/takdsp_rvv.S b/libavcodec/riscv/takdsp_rvv.S index 65c79e1aa9..816e765039 100644 --- a/libavcodec/riscv/takdsp_rvv.S +++ b/libavcodec/riscv/takdsp_rvv.S @@ -47,3 +47,20 @@ func ff_decorrelate_sr_rvv, zve32x bnez a2, 1b ret endfunc + +func ff_decorrelate_sm_rvv, zve32x +1: + vsetvli t0, a2, e32, m8, ta, ma + vle32.v v0, (a0) + sub a2, a2, t0 + vle32.v v8, (a1) + vsra.vi v16, v8, 1 + vsub.vv v0, v0, v16 + vse32.v v0, (a0) + sh2add a0, t0, a0 + vadd.vv v0, v0, v8 + vse32.v v0, (a1) + sh2add a1, t0, a1 + bnez a2, 1b + ret +endfunc -- 2.43.0