[FFmpeg-devel] [PATCH 14/18] avcodec/hevcdec: Don't allocate redundant HEVCContexts

[Andreas Rheinhardt]

Michael Niedermayer:
> On Sat, Jul 23, 2022 at 07:44:40AM +0200, Andreas Rheinhardt wrote:
>> Andreas Rheinhardt:
>>> Michael Niedermayer:
>>>> On Sat, Jul 02, 2022 at 08:32:06AM +0200, Andreas Rheinhardt wrote:
>>>>> Michael Niedermayer:
>>>>>> On Fri, Jul 01, 2022 at 12:29:45AM +0200, Andreas Rheinhardt wrote:
>>>>>>> The HEVC decoder has both HEVCContext and HEVCLocalContext
>>>>>>> structures. The latter is supposed to be the structure
>>>>>>> containing the per-slicethread state.
>>>>>>>
>>>>>>> Yet up until now that is not how it is handled in practice:
>>>>>>> Each HEVCLocalContext has a unique HEVCContext allocated for it
>>>>>>> and each of these coincides except in exactly one field: The
>>>>>>> corresponding HEVCLocalContext. This makes it possible to pass
>>>>>>> the HEVCContext everywhere where logically a HEVCLocalContext
>>>>>>> should be used. And up until recently, this is how it has been done.
>>>>>>>
>>>>>>> Yet the preceding patches changed this, making it possible
>>>>>>> to avoid allocating redundant HEVCContexts.
>>>>>>>
>>>>>>> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt at outlook.com>
>>>>>>> ---
>>>>>>>  libavcodec/hevcdec.c | 40 ++++++++++++++++------------------------
>>>>>>>  libavcodec/hevcdec.h |  2 --
>>>>>>>  2 files changed, 16 insertions(+), 26 deletions(-)
>>>>>>>
>>>>>>> diff --git a/libavcodec/hevcdec.c b/libavcodec/hevcdec.c
>>>>>>> index 9d1241f293..048fcc76b4 100644
>>>>>>> --- a/libavcodec/hevcdec.c
>>>>>>> +++ b/libavcodec/hevcdec.c
>>>>>>> @@ -2548,13 +2548,12 @@ static int hls_decode_entry_wpp(AVCodecContext *avctxt, void *hevc_lclist,
>>>>>>>  {
>>>>>>>      HEVCLocalContext *lc = ((HEVCLocalContext**)hevc_lclist)[self_id];
>>>>>>>      const HEVCContext *const s = lc->parent;
>>>>>>> -    HEVCContext *s1  = avctxt->priv_data;
>>>>>>> -    int ctb_size    = 1<< s1->ps.sps->log2_ctb_size;
>>>>>>> +    int ctb_size    = 1 << s->ps.sps->log2_ctb_size;
>>>>>>>      int more_data   = 1;
>>>>>>>      int ctb_row = job;
>>>>>>> -    int ctb_addr_rs = s1->sh.slice_ctb_addr_rs + ctb_row * ((s1->ps.sps->width + ctb_size - 1) >> s1->ps.sps->log2_ctb_size);
>>>>>>> -    int ctb_addr_ts = s1->ps.pps->ctb_addr_rs_to_ts[ctb_addr_rs];
>>>>>>> -    int thread = ctb_row % s1->threads_number;
>>>>>>> +    int ctb_addr_rs = s->sh.slice_ctb_addr_rs + ctb_row * ((s->ps.sps->width + ctb_size - 1) >> s->ps.sps->log2_ctb_size);
>>>>>>> +    int ctb_addr_ts = s->ps.pps->ctb_addr_rs_to_ts[ctb_addr_rs];
>>>>>>> +    int thread = ctb_row % s->threads_number;
>>>>>>>      int ret;
>>>>>>>  
>>>>>>>      if(ctb_row) {
>>>>>>> @@ -2572,7 +2571,7 @@ static int hls_decode_entry_wpp(AVCodecContext *avctxt, void *hevc_lclist,
>>>>>>>  
>>>>>>>          ff_thread_await_progress2(s->avctx, ctb_row, thread, SHIFT_CTB_WPP);
>>>>>>>  
>>>>>>> -        if (atomic_load(&s1->wpp_err)) {
>>>>>>> +        if (atomic_load(&s->wpp_err)) {
>>>>>>>              ff_thread_report_progress2(s->avctx, ctb_row , thread, SHIFT_CTB_WPP);
>>>>>>
>>>>>> the consts in "const HEVCContext *const " make clang version 6.0.0-1ubuntu2 unhappy
>>>>>> (this was building shared libs)
>>>>>>
>>>>>>
>>>>>> CC	libavcodec/hevcdec.o
>>>>>> src/libavcodec/hevcdec.c:2574:13: error: address argument to atomic operation must be a pointer to non-const _Atomic type ('const atomic_int *' (aka 'const _Atomic(int) *') invalid)
>>>>>>         if (atomic_load(&s->wpp_err)) {
>>>>>>             ^           ~~~~~~~~~~~
>>>>>> /usr/lib/llvm-6.0/lib/clang/6.0.0/include/stdatomic.h:134:29: note: expanded from macro 'atomic_load'
>>>>>> #define atomic_load(object) __c11_atomic_load(object, __ATOMIC_SEQ_CST)
>>>>>>                             ^                 ~~~~~~
>>>>>> 1 error generated.
>>>>>> src/ffbuild/common.mak:81: recipe for target 'libavcodec/hevcdec.o' failed
>>>>>> make: *** [libavcodec/hevcdec.o] Error 1
>>>>>>
>>>>>> thx
>>>>>>
>>>>>
>>>>> Thanks for testing this. atomic_load is indeed declared without const in
>>>>> 7.17.7.2:
>>>>>
>>>>> C atomic_load(volatile A *object);
>>>>>
>>>>> Upon reflection this makes sense, because if atomics are implemented via
>>>>> mutexes, even a read may involve a preceding write. So I'll cast const
>>>>> away here, too, and add a comment. (It works when casting const away,
>>>>> doesn't it?)
>>>>
>>>> This doesnt feel "right". These pointers should not be coming from a const
>>>> if they are written to
>>>>
>>>
>>> The HEVCContext is not const because the underlying object is const; the
>>> HEVCContext is const when accessed from any part of the code that may be
>>> run from slice threads, because if a slice thread modifies it, you have
>>> a data race in case any of the other slice threads reads this field or
>>> modifies it itself. But this is by definition not true for atomic
>>> operations, so casting const away for them is fine.
>>>
>>>> The compiler accepts it with an explicit cast though. With an implicit cast
>>>> it produces a warning
>>>>
>>
>> Did the above explanation satisfy you? Or do you want something else?
> 
> sure, ok
> 
> [...]
> 

Good to hear. This patchset (namely patch 11/18: "avcodec/hevcpred: Pass
HEVCLocalContext when slice-threading") includes modifications to mips
code that I created blindly. Can you please test it? Here is a branch of
this rebased on top of current git master:
https://github.com/mkver/FFmpeg/commits/hevc_wpp
(Said branch actually contains a bit of further work which also modifies
mips code (in particular,
https://github.com/mkver/FFmpeg/commit/cf441e559b8d4bf2c05c29483ccf49e82fc6b863
does so); you may also test this.)

- Andreas