[x265] [PATCH] p2s-sve.S: Fix filterPixelToShort_48x64 for longer SVE vectors

[George Steed]

Ping :)

On Tue, Mar 11, 2025 at 04:11:17PM +0000, George Steed wrote:
> The existing filterPixelToShort_48x64 high bit-depth code for vectors of
> at least 256-bits uses predication to ensure that only the low 256 bits
> of the vector are operated on, however the address arithmetic is
> performed relative to the vector length so for vectors of 512-bits or
> longer this is incorrect. Since we are operating on the fixed low 256
> bits of the vectors here, fix the code by hard-coding the address offset
> to multiples of 32 bytes.
> ---
>  source/common/aarch64/p2s-sve.S | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/source/common/aarch64/p2s-sve.S b/source/common/aarch64/p2s-sve.S
> index 85bb14b3d..11e63ddab 100644
> --- a/source/common/aarch64/p2s-sve.S
> +++ b/source/common/aarch64/p2s-sve.S
> @@ -401,10 +401,12 @@ function PFX(filterPixelToShort_48x64_sve)
>      ret
>  .vl_gt_16_filterPixelToShort_high_48x64:
>      ptrue           p0.h, vl16
> +    mov             x4, #16
> +    mov             x5, #32
>  .rept 64
>      ld1h            {z0.h}, p0/z, [x0]
> -    ld1h            {z1.h}, p0/z, [x0, #1, mul vl]
> -    ld1h            {z2.h}, p0/z, [x0, #2, mul vl]
> +    ld1h            {z1.h}, p0/z, [x0, x4, lsl #1]
> +    ld1h            {z2.h}, p0/z, [x0, x5, lsl #1]
>      add             x0, x0, x1
>      lsl             z0.h, p0/m, z0.h, #P2S_SHIFT
>      lsl             z1.h, p0/m, z1.h, #P2S_SHIFT
> @@ -413,8 +415,8 @@ function PFX(filterPixelToShort_48x64_sve)
>      add             z1.h, p0/m, z1.h, z31.h
>      add             z2.h, p0/m, z2.h, z31.h
>      st1h            {z0.h}, p0, [x2]
> -    st1h            {z1.h}, p0, [x2, #1, mul vl]
> -    st1h            {z2.h}, p0, [x2, #2, mul vl]
> +    st1h            {z1.h}, p0, [x2, x4, lsl #1]
> +    st1h            {z2.h}, p0, [x2, x5, lsl #1]
>      add             x2, x2, x3
>  .endr
>      ret
> -- 
> 2.34.1
> 


> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel

-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-p2s-sve.S-Fix-filterPixelToShort_48x64-for-longer-SV.patch
Type: text/x-diff
Size: 2021 bytes
Desc: not available
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20250416/9b4e95bc/attachment.patch>