diff mbox series

[FFmpeg-devel] lavu/float_dsp: rework RISC-V V scalar product

Message ID 20230717172557.15848-1-remi@remlab.net
State New
Headers show
Series [FFmpeg-devel] lavu/float_dsp: rework RISC-V V scalar product | expand

Checks

Context Check Description
yinshiyou/make_loongarch64 success Make finished
yinshiyou/make_fate_loongarch64 success Make fate finished
andriy/make_x86 success Make finished
andriy/make_fate_x86 success Make fate finished

Commit Message

Rémi Denis-Courmont July 17, 2023, 5:25 p.m. UTC 
1) Take the reductive sum out of the loop,
   leaving a regular vector addition in the loop.
2) Merge the addition and the multiplication.
3) Unroll.

Before:
scalarproduct_float_rvv_f32: 832.5

After:
scalarproduct_float_rvv_f32: 275.2
---
 libavutil/riscv/float_dsp_rvv.S | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

Comments

Rémi Denis-Courmont July 17, 2023, 5:51 p.m. UTC | #1 
Le maanantaina 17. heinäkuuta 2023, 20.25.57 EEST Rémi Denis-Courmont a écrit 
:
> 1) Take the reductive sum out of the loop,
>    leaving a regular vector addition in the loop.
> 2) Merge the addition and the multiplication.
> 3) Unroll.
> 
> Before:
> scalarproduct_float_rvv_f32: 832.5
> 
> After:
> scalarproduct_float_rvv_f32: 275.2
> ---
>  libavutil/riscv/float_dsp_rvv.S | 13 +++++++------
>  1 file changed, 7 insertions(+), 6 deletions(-)
> 
> diff --git a/libavutil/riscv/float_dsp_rvv.S
> b/libavutil/riscv/float_dsp_rvv.S index 77961b7387..c80fde2f7e 100644
> --- a/libavutil/riscv/float_dsp_rvv.S
> +++ b/libavutil/riscv/float_dsp_rvv.S
> @@ -166,20 +166,21 @@ endfunc
> 
>  // a0 = (a0).(a1) [0..a2-1]
>  func ff_scalarproduct_float_rvv, zve32f
> -        vsetivli     zero, 1, e32, m1, ta, ma
> -        vmv.s.x      v8, zero
> +        vsetvli      t0, zero, e32, m8, ta, ma
> +        vmv.v.x      v8, zero
> +        vmv.s.x      v0, zero
>  1:
> -        vsetvli      t0, a2, e32, m1, ta, ma
> +        vsetvli      t0, a2, e32, m8, tu, ma
>          vle32.v      v16, (a0)
>          sub          a2, a2, t0
>          vle32.v      v24, (a1)
>          sh2add       a0, t0, a0
> -        vfmul.vv     v16, v16, v24
> +        vfmacc.vv    v8, v16, v24
>          sh2add       a1, t0, a1
> -        vfredusum.vs v8, v16, v8
>          bnez         a2, 1b
> 
> -        vfmv.f.s fa0, v8
> +        vfredusum.vs v0, v8, v0

Missing vsetvli; this won't work unless the input vector size is a multiple of 
the hardware vector length.

> +        vfmv.f.s     fa0, v0
>  NOHWF   fmv.x.w  a0, fa0
>          ret
>  endfunc
diff mbox series

Patch

diff --git a/libavutil/riscv/float_dsp_rvv.S b/libavutil/riscv/float_dsp_rvv.S
index 77961b7387..c80fde2f7e 100644
--- a/libavutil/riscv/float_dsp_rvv.S
+++ b/libavutil/riscv/float_dsp_rvv.S
@@ -166,20 +166,21 @@  endfunc
 
 // a0 = (a0).(a1) [0..a2-1]
 func ff_scalarproduct_float_rvv, zve32f
-        vsetivli     zero, 1, e32, m1, ta, ma
-        vmv.s.x      v8, zero
+        vsetvli      t0, zero, e32, m8, ta, ma
+        vmv.v.x      v8, zero
+        vmv.s.x      v0, zero
 1:
-        vsetvli      t0, a2, e32, m1, ta, ma
+        vsetvli      t0, a2, e32, m8, tu, ma
         vle32.v      v16, (a0)
         sub          a2, a2, t0
         vle32.v      v24, (a1)
         sh2add       a0, t0, a0
-        vfmul.vv     v16, v16, v24
+        vfmacc.vv    v8, v16, v24
         sh2add       a1, t0, a1
-        vfredusum.vs v8, v16, v8
         bnez         a2, 1b
 
-        vfmv.f.s fa0, v8
+        vfredusum.vs v0, v8, v0
+        vfmv.f.s     fa0, v0
 NOHWF   fmv.x.w  a0, fa0
         ret
 endfunc