[FFmpeg-devel] libavcodec/als: remove check for predictor order of a block

[Thilo Borgmann]

Am 13.11.17 um 20:02 schrieb Umair Khan:
> Hi,
> 
> On Mon, Nov 13, 2017 at 11:06 PM, Thilo Borgmann <thilo.borgmann at mail.de> wrote:
>> Hi,
>>
>>> On Mon, Nov 13, 2017 at 1:09 AM, Carl Eugen Hoyos <ceffmpeg at gmail.com> wrote:
>>>> 2017-11-12 20:30 GMT+01:00 Umair Khan <omerjerk at gmail.com>:
>>>>> Hi,
>>>>>
>>>>> On Mon, Nov 13, 2017 at 12:45 AM, Carl Eugen Hoyos <ceffmpeg at gmail.com> wrote:
>>>>>> 2017-11-12 20:05 GMT+01:00 Umair Khan <omerjerk at gmail.com>:
>>>>>>
>>>>>>> The attached patch fixes the address sanitizer issue.
>>>>>>
>>>>>> Breaks compilation here, how did you test?
>>>>>>
>>>>>> libavcodec/alsdec.c: In function ‘decode_var_block_data’:
>>>>>> libavcodec/alsdec.c:938:7: error: expected ‘}’ before ‘else’
>>>>>
>>>>> Sorry for the faulty patch. Here is the fixed one.
>>>>
>>>> The commit message of your patch is:
>>>> libavcodec/als: fix address sanitization error in decoder
>>>>
>>>> Is there an error in current FFmpeg git head that asan
>>>> shows? If not, the commit message makes no sense.
>>>>
>>>> I believe you should send two patches that are meant
>>>> to be committed together, one of them fixing ticket #6297.
>>>
>>> This is the complete patchset.
>>
>> I need some days to find time to test this, earliest during the weekend I fear...
>>
>> What happens for
>> block_length < residual_index < opt_order?
> 
> I didn't really understand this case. What's residual_index? Can you
> point to the source exactly?
> 
> As far as the case where opt_order is more than block_length, my
> second patch handles that case only. The file which Michael sent was
> having asan issues because of the case when block_length < opt_order.
> 
>> Another way of asking would be, where is the second loop from specs page 30 for that case?
>> (ISO/IEC 14496)
> 
> The second loop is just converting parcor to lpc coefficients which is
> done here - https://github.com/FFmpeg/FFmpeg/blob/master/libavcodec/alsdec.c#L935
> 
>> I think what puzzles CE is, that the problematic if() from the other patches is still untouched by your patch. So how could this be a valid solution even if your patch would actually improve the prediction part...
>> And I wonder the same ;)
> 
> As said, it is valid to have opt_order greater than block_length.
> However, the decoder loop needs to be checked because we won't predict
> values more than the length of the block i.e., block_length. We use
> last K (prediction order, opt_order) values to predict the original K
> values of the current block.
> 
>> Did you run FATE with your patch applied? I assume a big difference in output at the first glance (means FATE aks the conformance files should fail...)
> 
> Yes. I did run FATE. It passes perfectly.
> 
>> Thanks for driving this forward anyway :)
> 
> I think the two patches fix the issues completely. I don't see any
> harm in applying this patchset. :)

Which second patch exactly do you want to be applied along with this one?

-Thilo