[FFmpeg-devel] [PATCH 2/2] sws/rgb2rgb: fix unaligned accesses in R-V V YUYV to I422p

[Rémi Denis-Courmont]

Le torstaina 9. marraskuuta 2023, 20.34.53 EET Rémi Denis-Courmont a écrit :
> In my personal opinion, we should not need to support unaligned YUY2
> pixel maps. They should always be aligned to at least 32 bits, and the
> current code assumes just 16 bits. However checkasm does test for
> unaligned input bitmaps. QEMU accepts it, but real hardware dose not.
> 
> In this particular case, we can at the same time improve performance and
> handle unaligned inputs, so do just that.
> 
> uyvytoyuv422_c:      104060.0
> uyvytoyuv422_rvv_i32: 25284.0 (before)
> uyvytoyuv422_rvv_i32: 20148.2 (after)
> ---
>  libswscale/riscv/rgb2rgb_rvv.S | 45 +++++++++++++++++-----------------
>  1 file changed, 23 insertions(+), 22 deletions(-)
> 
> diff --git a/libswscale/riscv/rgb2rgb_rvv.S b/libswscale/riscv/rgb2rgb_rvv.S
> index 172f5918dc..716948dc82 100644
> --- a/libswscale/riscv/rgb2rgb_rvv.S
> +++ b/libswscale/riscv/rgb2rgb_rvv.S
> @@ -126,32 +126,33 @@ func ff_deinterleave_bytes_rvv, zve32x
>          ret
>  endfunc
> 
> -.macro yuy2_to_i422p y_shift
> -        slli    t4, a4, 1 // pixel width -> (source) byte width
> +.macro yuy2_to_i422p luma, chroma
> +        srai    t4, a4, 1 // pixel width -> chroma width
>          lw      t6, (sp)
> +        slli    t5, a4, 1 // pixel width -> (source) byte width
>          sub     a6, a6, a4
> -        srai    a4, a4, 1 // pixel width -> chroma width
> -        sub     a7, a7, a4
> -        sub     t6, t6, t4
> +        sub     a7, a7, t4
> +        sub     t6, t6, t5
>  1:
>          mv      t4, a4
>          addi    a5, a5, -1
>  2:
> -        vsetvli    t5, t4, e8, m2, ta, ma
> -        vlseg2e16.v v16, (a3)
> -        sub        t4, t4, t5
> -        vnsrl.wi   v24, v16, \y_shift // Y0
> -        sh2add     a3, t5, a3
> -        vnsrl.wi   v26, v20, \y_shift // Y1
> -        vnsrl.wi   v28, v16, 8 - \y_shift // U
> -        vnsrl.wi   v30, v20, 8 - \y_shift // V
> -        vsseg2e8.v v24, (a0)
> -        sh1add     a0, t5, a0
> -        vse8.v     v28, (a1)
> -        add        a1, t5, a1
> -        vse8.v     v30, (a2)
> -        add        a2, t5, a2
> -        bnez       t4, 2b
> +        vsetvli t5, t4, e8, m4, ta, ma
> +        vlseg2e8.v v16, (a3)
> +        srli    t1, t5, 1

Due to the specifications of RVV, we cannot assume that on the penultimate loop 
iteration, the selected vector length (t5) will be even. If it is odd, then 
this won't work. The test hardware always picks the maximum possible, which is 
always even, but...

This needs special handling :(

> +        vsetvli zero, t1, e8, m2, ta, ma
> +        vnsrl.wi   v24, \chroma, 0 // U
> +        sub     t4, t4, t5
> +        vnsrl.wi   v28, \chroma, 8 // V
> +        sh1add  a3, t5, a3
> +        vse8.v  v24, (a1)
> +        add     a1, t1, a1
> +        vse8.v  v28, (a2)
> +        add     a2, t1, a2
> +        vsetvli zero, t5, e8, m4, ta, ma
> +        vse8.v  \luma, (a0)
> +        add     a0, t5, a0
> +        bnez    t4, 2b
> 
>          add     a3, a3, t6
>          add     a0, a0, a6
> @@ -163,9 +164,9 @@ endfunc
>  .endm
> 
>  func ff_uyvytoyuv422_rvv, zve32x
> -        yuy2_to_i422p 8
> +        yuy2_to_i422p v20, v16
>  endfunc
> 
>  func ff_yuyvtoyuv422_rvv, zve32x
> -        yuy2_to_i422p 0
> +        yuy2_to_i422p v16, v20
>  endfunc


-- 
Rémi Denis-Courmont
http://www.remlab.net/