[pulseaudio-discuss] [PATCH] bluetooth: Fix rendering a2dp data

[Luiz Augusto von Dentz]

Hi,

On Tue, Dec 14, 2010 at 11:10 AM, Maarten Lankhorst
<m.b.lankhorst at gmail.com> wrote:
>  Hi Luiz,
>
> Op 13-12-10 09:55, Luiz Augusto von Dentz schreef:
>>
>> Hi,
>>
>> On Sat, Dec 11, 2010 at 1:05 AM, Maarten Lankhorst
>> <m.b.lankhorst at gmail.com>  wrote:
>>>
>>> makes my android phone slightly happier
>>> ---
>>>  src/modules/bluetooth/module-bluetooth-device.c |   11 ++++++++---
>>>  1 files changed, 8 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/src/modules/bluetooth/module-bluetooth-device.c
>>> b/src/modules/bluetooth/module-bluetooth-device.c
>>> index 6d31c1e..8664001 100644
>>> --- a/src/modules/bluetooth/module-bluetooth-device.c
>>> +++ b/src/modules/bluetooth/module-bluetooth-device.c
>>> @@ -1387,7 +1387,7 @@ static int a2dp_process_push(struct userdata *u) {
>>>     pa_assert(u->source);
>>>     pa_assert(u->read_smoother);
>>>
>>> -    memchunk.memblock = pa_memblock_new(u->core->mempool,
>>> u->block_size);
>>> +    memchunk.memblock = pa_memblock_new(u->core->mempool, u->block_size
>>> * 2);
>>>     memchunk.index = memchunk.length = 0;
>>
>> Im not sure how this would help, we decode sbc frame by frame so
>> having twice as much memory doesn't really make any sense, there could
>> be that we should decode the entire buffer read, but that would take
>> much more memory. What could be the real problem is that the source is
>> changing the bitpool but that shouldn't be a problem since the
>> block_size is supposed to be calculated from maximum bitpool range so
>> the buffer will always be big enough.
>
> Well it seems the block_size is not fixed on my phone, sometimes it is 1
> sample more, sometimes 2, the net effect is 1 or 2 frames stay undecoded
> leading to massive underruns.
>
> I don't think u->block_size has to be fixed per se, a2dp->code_size seems to
> be though.

Again if we want to be able to support devices that changes bitpool we
have to recalculate the block_size on every frame (if bitpool has
really changed), doubling the buffer is not the right fix, sorry.

>>>
>>>     for (;;) {
>>> @@ -1442,7 +1442,8 @@ static int a2dp_process_push(struct userdata *u) {
>>>         to_decode = l - sizeof(*header) - sizeof(*payload);
>>>
>>>         d = pa_memblock_acquire(memchunk.memblock);
>>> -        to_write = memchunk.length =
>>> pa_memblock_get_length(memchunk.memblock);
>>> +        to_write = pa_memblock_get_length(memchunk.memblock);
>>> +        memchunk.length = 0;
>>>
>>>         while (PA_LIKELY(to_decode>  0&&  to_write>  0)) {
>>>             size_t written;
>>> @@ -1464,7 +1465,7 @@ static int a2dp_process_push(struct userdata *u) {
>>>  /*             pa_log_debug("SBC: frame_length: %lu; codesize: %lu",
>>> (unsigned long) a2dp->frame_length, (unsigned long) a2dp->codesize); */
>>>
>>>             pa_assert_fp((size_t) decoded<= to_decode);
>>> -            pa_assert_fp((size_t) decoded == a2dp->frame_length);
>>> +            pa_assert_fp((size_t) decoded<= a2dp->frame_length);
>>
>> This is the real problem, either we do this if the bitpool is changing
>> or we have to call sbc_reinit in each frame.
>
> Well, a2dp->frame_length was calculated based on the maximum bitpool size.
> The frames decoded have a smaller bitpool size. a2dp->codesize is still the
> same though, so I don't think this change matters much.

It was you that suggested this changes, so when you say it doesn't
matter it makes me wonder why this is part of the patch then, Btw,
changing bitpool changes the frame length and it actually does assert,
I tried it myself with this patch to source:

http://gitorious.org/pulseaudio/mainline/commit/9b938f8f8f40772e626b3e71fa4d7b7cbd5de5e7

>>>
>>>             pa_assert_fp((size_t) written<= to_write);
>>>             pa_assert_fp((size_t) written == a2dp->codesize);
>>> @@ -1474,10 +1475,14 @@ static int a2dp_process_push(struct userdata *u)
>>> {
>>>
>>>             d = (uint8_t*) d + written;
>>>             to_write -= written;
>>> +            memchunk.length += written;
>>>
>>>             frame_count++;
>>>         }
>>>
>>> +        if (to_decode)
>>> +            pa_log_error("SBC: %lu bytes not decoded\n", to_decode);
>>> +
>>>         pa_memblock_release(memchunk.memblock);
>>
>> Actually I think we should be solving this in libsbc, both sbc_encode
>> and sbc_decode should be checking if the bitpool has changed and
>> reinit automatically, otherwise we gonna get performance penalty doing
>> sbc_reinit for every frame.
>
> I don't think the bitpool changed, the code just assumes the maximum bitpool
> size, while my phone sends data at a lower one.

Well what other bitpool would you use instead? If we don't have any
frame to parse obviously we need to start with the biggest possible
one, it the only possible way to avoid decode errors due to too small
buffer to decode.

-- 
Luiz Augusto von Dentz
Computer Engineer