[v2,0/2] media: verisilicon: fix reference padding issue

Message

Benjamin Gaignard Dec. 12, 2024, 3:43 p.m. UTC

Hantro hardware is splited into two parts: the decoder which produce
tiled pixels formats and the post-processor which produce raster pixels
formats.
When post-processor is used the selected pixels format may not have the
some padding/stride than the decoder output pixels format. This led to
miscomputing chroma and motion vectors offsets so data are overlapping.
This series introduce a reference pixels format that decoder use to
compute the various needed offsets and size.

With this series Fluster for VP9 tests is now 207/305 vs 145/305.
HEVC test score isn't impacted by these patches (still 141/147).

Version 2:
- rebased on top of media-commiter/next
- Add reviewed-by tags

Benjamin Gaignard (2):
  media: verisilicon: Store reference frames pixels format
  media: verisilicon: Fix IMX8 native pixels format steps values

 drivers/media/platform/verisilicon/hantro.h   |  2 ++
 .../media/platform/verisilicon/hantro_g2.c    |  2 +-
 .../platform/verisilicon/hantro_postproc.c    | 32 +++++++------------
 .../media/platform/verisilicon/hantro_v4l2.c  | 21 ++++++++++++
 .../media/platform/verisilicon/imx8m_vpu_hw.c | 10 +++---
 5 files changed, 40 insertions(+), 27 deletions(-)

Comments

Fabio Estevam Dec. 12, 2024, 4:10 p.m. UTC | #1

On Thu, Dec 12, 2024 at 12:43 PM Benjamin Gaignard
<[e-mail]> wrote:
>
> Hantro hardware is splited into two parts: the decoder which produce
> tiled pixels formats and the post-processor which produce raster pixels
> formats.
> When post-processor is used the selected pixels format may not have the
> some padding/stride than the decoder output pixels format. This led to
> miscomputing chroma and motion vectors offsets so data are overlapping.
> This series introduce a reference pixels format that decoder use to
> compute the various needed offsets and size.
>
> With this series Fluster for VP9 tests is now 207/305 vs 145/305.
> HEVC test score isn't impacted by these patches (still 141/147).
>
> Version 2:
> - rebased on top of media-commiter/next
> - Add reviewed-by tags
>
> Benjamin Gaignard (2):
>   media: verisilicon: Store reference frames pixels format
>   media: verisilicon: Fix IMX8 native pixels format steps values

Do both patches deserve a Fixes tag?

Benjamin Gaignard Dec. 12, 2024, 4:15 p.m. UTC | #2

Le 12/12/2024 à 17:10, Fabio Estevam a écrit :
> On Thu, Dec 12, 2024 at 12:43 PM Benjamin Gaignard
> <[e-mail]> wrote:
>> Hantro hardware is splited into two parts: the decoder which produce
>> tiled pixels formats and the post-processor which produce raster pixels
>> formats.
>> When post-processor is used the selected pixels format may not have the
>> some padding/stride than the decoder output pixels format. This led to
>> miscomputing chroma and motion vectors offsets so data are overlapping.
>> This series introduce a reference pixels format that decoder use to
>> compute the various needed offsets and size.
>>
>> With this series Fluster for VP9 tests is now 207/305 vs 145/305.
>> HEVC test score isn't impacted by these patches (still 141/147).
>>
>> Version 2:
>> - rebased on top of media-commiter/next
>> - Add reviewed-by tags
>>
>> Benjamin Gaignard (2):
>>    media: verisilicon: Store reference frames pixels format
>>    media: verisilicon: Fix IMX8 native pixels format steps values
> Do both patches deserve a Fixes tag?

There is dependency in around hevc compress feature in post-proc code.
I don't think it could be applied on stable kernel