[FFmpeg-devel] [PATCH] avcodec/h264_mb: Fix buffer stride for arm

[Michael Niedermayer]

On Tue, Apr 08, 2025 at 07:22:32PM +0800, Bin Peng wrote:
> When decoding a bitstream with weighted-bipred enabled,
> the results on ARM and x86 platforms may differ.
> 
> The reason for the inconsistency is that the value of
> STRIDE_ALIGN differs between platforms. And STRIDE_ALIGN
> is set to the buffer stride of temporary buffers for U
> and V components in mc_part_weighted.
> 
> If the buffer stride is 32 or 64 (as on x86 platforms),
> the U and V pixels can be interleaved row by row without
> overlapping, resulting in correct output.
> However, on ARM platforms where the stride is 16,
> the V component will overwrite part of the U component's pixels,
> leading to incorrect predicted pixels.
> 
> The bug can be reproduced by the following bitstream.
> 
> https://trac.ffmpeg.org/attachment/ticket/11357/inter_weighted_bipred2.264
> 
> And the ref/fate file is also added in this patch.
> 
> Fixes: ticket 11357
> Signed-off-by: Bin Peng <pengbin at visionular.com>
> ---
>  libavcodec/utils.c                            |  4 ++-

>  .../h264-conformance-weighted_bipred2.264     | 31 +++++++++++++++++++

is this supposed to be a fate test ?
if so theres something missing also the file extension is a bit misleading
and tests covering more cases is always welcome


>  2 files changed, 34 insertions(+), 1 deletion(-)
>  create mode 100644 tests/ref/fate/h264-conformance-weighted_bipred2.264
> 
> diff --git a/libavcodec/utils.c b/libavcodec/utils.c
> index 90867ed6b1..5d15f5c8fa 100644
> --- a/libavcodec/utils.c
> +++ b/libavcodec/utils.c
> @@ -144,6 +144,7 @@ void avcodec_align_dimensions2(AVCodecContext *s, int *width, int *height,
>      int i;
>      int w_align = 1;
>      int h_align = 1;
> +    int stride_align = STRIDE_ALIGN;
>      AVPixFmtDescriptor const *desc = av_pix_fmt_desc_get(s->pix_fmt);
>  
>      if (desc) {
> @@ -339,13 +340,14 @@ void avcodec_align_dimensions2(AVCodecContext *s, int *width, int *height,
>          // increasing witdth ensure that the temporary area is large enough,
>          // the next rounded up width is 32
>          *width = FFMAX(*width, 32);
> +        stride_align = FFMAX(stride_align, 32);

does the following fix this too ?

diff --git a/libavcodec/h264_mb.c b/libavcodec/h264_mb.c
index 6083f7ad84f..0d6562b5830 100644
--- a/libavcodec/h264_mb.c
+++ b/libavcodec/h264_mb.c
@@ -407,7 +407,7 @@ static av_always_inline void mc_part_weighted(const H264Context *h, H264SliceCon
         /* don't optimize for luma-only case, since B-frames usually
          * use implicit weights => chroma too. */
         uint8_t *tmp_cb = sl->bipred_scratchpad;
-        uint8_t *tmp_cr = sl->bipred_scratchpad + (16 << pixel_shift);
+        uint8_t *tmp_cr = sl->bipred_scratchpad + (8 << pixel_shift + (chroma_idc == 3));
         uint8_t *tmp_y  = sl->bipred_scratchpad + 16 * sl->mb_uvlinesize;
         int refn0       = sl->ref_cache[0][scan8[n]];
         int refn1       = sl->ref_cache[1][scan8[n]];

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Never trust a computer, one day, it may think you are the virus. -- Compn
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 195 bytes
Desc: not available
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20250416/793ba490/attachment.sig>