Re: [PATCH v7 2/4] media: rockchip: Add JPEG decoder driver

From: Sergei Khomenkov

Date: Thu Nov 19 2026 - 15:22:04 EST


Hi Sascha,

On 2026-11-18 14:27, Sascha Hauer wrote:
> I haven't been affected by this issue lately, but saw it when the series
> was still based on the hantro driver. Yes, I'll change to 3 bytes per
> pixel next round and make that the floor of sizeimage. Would be nice to
> have a proper solution for this rather than either wasting tons of
> memory or risk having too small buffers.

Thanks. A correction to my last mail: 3 bytes per pixel is not a worst
case for 4:2:0, despite what tj3JPEGBufSize() suggests. Since then I
built 4:2:0 test patterns that push the quality 100 scan as far as I
could: libjpeg-turbo writes 3.35 bytes per pixel of scan data, byte
stuffing included, for one of them (640x480), and the Hantro H1
encoder on the RK3026 3.14. At quality 99 the largest I found was 3.08
(libjpeg-turbo) and 2.88 (H1). The patterns, 96x48 and 640x480 JPEGs,
the same frames as raw NV12 and a short script that makes them at any
size, are at https://homenkow.ru/kernel/vpu/

With the Annex K Huffman tables and 8-bit samples, a baseline 4:2:0
scan is at most about 3.3 bytes per pixel before byte stuffing and 4.9
after it, at quality 100 (lower qualities are smaller). Other tables
give a higher bound, and APPn/COM segments add any amount on top, so
for a decoder any fixed factor is a trade-off.

So 3 bytes per pixel covers what the H1 wrote up to quality 99 and most
of its quality 100 frames, but libjpeg-turbo goes over it from quality
99 up. If that trade-off is fine for you, I have no objection; I only
would not call it the worst case in the commit message or a comment.

Sergei