Re: [PATCH v8 3/3] media: cedrus: Add HEVC/H.265 decoding support

[Paul Kocialkowski]

Hi,

On Tue 22 Oct 19, 15:37, Hans Verkuil wrote:
> On 10/22/19 3:17 PM, Paul Kocialkowski wrote:
> > Hi again,
> > 
> > On Tue 22 Oct 19, 14:40, Paul Kocialkowski wrote:
> >> Hi Mauro and thanks for the review,
> >>
> >> On Thu 17 Oct 19, 09:57, Mauro Carvalho Chehab wrote:
> >>> Em Fri, 27 Sep 2019 16:34:11 +0200
> >>> Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx> escreveu:
> >>>
> >>>> This introduces support for HEVC/H.265 to the Cedrus VPU driver, with
> >>>> both uni-directional and bi-directional prediction modes supported.
> >>>>
> >>>> Field-coded (interlaced) pictures, custom quantization matrices and
> >>>> 10-bit output are not supported at this point.
> >>>>
> >>>> Signed-off-by: Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx>
> >>>> ---
> >>>
> >>> ...
> >>>
> >>>> +		unsigned int ctb_size_luma =
> >>>> +			1 << log2_max_luma_coding_block_size;
> >>>
> >>> Shifts like this is a little scary. "1" constant is signed. So, if
> >>> log2_max_luma_coding_block_size is 31, the above logic has undefined
> >>> behavior. Different archs and C compilers may handle it on different
> >>> ways.
> >>
> >> I wasn't aware that it was the case, thanks for bringing this to light!
> >> I'll make it 1UL then.
> >>
> >>>> +#define VE_DEC_H265_LOW_ADDR_PRIMARY_CHROMA(a) \
> >>>> +	(((a) << 24) & GENMASK(31, 24))
> >>>
> >>> Same applies here and on other similar macros. You need to enforce
> >>> (a) to be unsigned, as otherwise the behavior is undefined.
> >>>
> >>> Btw, this is a recurrent pattern on this file. I would define a
> >>> macro, e. g. something like:
> >>>
> >>> 	#define MASK_BITS_AND_SHIFT(v, high, low) \
> >>> 		((UL(v) << low) & GENMASK(high, low))
> >>>
> >>> And use it for all similar patterns here.
> >>
> >> Sounds good! I find that the reverse wording (SHIFT_AND_MASK_BITS) would be
> >> a bit more explicit since the shift happens prior to the mask.
> > 
> > Apparently the UL(v) macro just appends UL to v in preprocessor, so it won't
> > work with anything else than direct integers.
> > 
> > I'll replace it with a (unsigned long) cast, that seems to do the job.
> 
> Shouldn't that be a (u32) cast? Since this is used with 32 bit registers?

This would work for cedrus, but I think that what Mauro had in mind was to
migrate this macro to linux/bits.h, where everthing else (including GENMASK)
is apparently defined in terms of unsigned long and not types with explicit
numbers of bits. So I find it more consistent to go with unsigned long.

In our case, 64-bit platforms that use cedrus would calculate the macro on
64 bits and use it in 32-bit variables. Since we're never masking beyond the
lower 32 bits, I don't see how things could go wrong and the situation looks
fairly similar to the use of GENMASK in similar conditions.

Does that sound right to you or am I missing something here?

Cheers,

Paul

> Regards,
> 
> 	Hans
> 
> > 
> > Cheers,
> > 
> > Paul
> > 
> >> Also we probably need to have parenthesis around "low", right?
> >>
> >>> The best would be to include such macro at linux/bits.h, although some
> >>> upstream discussion is required.
> >>>
> >>> So, for now, let's add it at this header file, but work upstream
> >>> to have it merged there.
> >>
> >> Understood, I'll include it in that header for now and send a separate patch
> >> for inclusion in linux/bits.h (apparently the preprocessor doesn't care about
> >> redefinitions so we can just remove the cedrus fashion once the common one is
> >> in).
> >>
> >> What do you think?
> >>
> >> Cheers,
> >>
> >> Paul
> > 
> > 
> > 
> 

-- 
Paul Kocialkowski, Bootlin
Embedded Linux and kernel engineering
https://bootlin.com

Attachment: signature.asc
Description: PGP signature