[FFmpeg-devel] [PATCH] vulkan_hevc: switch from a buffer pool to a simple malloc

[Andreas Rheinhardt]

Lynne:
> Sep 15, 2023, 03:00 by andreas.rheinhardt at outlook.com:
> 
>> Lynne:
>>
>>> -static int get_data_set_buf(FFVulkanDecodeContext *s, AVBufferRef **data_buf,
>>> -                            int nb_vps, AVBufferRef * const vps_list[HEVC_MAX_VPS_COUNT])
>>> +static int alloc_hevc_header_structs(FFVulkanDecodeContext *s,
>>> +                                     int nb_vps,
>>> +                                     AVBufferRef * const vps_list[HEVC_MAX_VPS_COUNT])
>>>  {
>>> -    uint8_t *data_ptr;
>>> +    uintptr_t data_ptr;
>>>  HEVCHeaderSet *hdr;
>>>  
>>>  size_t base_size = sizeof(StdVideoH265SequenceParameterSet)*HEVC_MAX_SPS_COUNT +
>>> @@ -93,22 +94,18 @@ static int get_data_set_buf(FFVulkanDecodeContext *s, AVBufferRef **data_buf,
>>>  buf_size += sizeof(HEVCHeaderVPSSet)*vps->vps_num_hrd_parameters;
>>>  }
>>>  
>>> -    if (buf_size > s->tmp_pool_ele_size) {
>>> -        av_buffer_pool_uninit(&s->tmp_pool);
>>> -        s->tmp_pool_ele_size = 0;
>>> -        s->tmp_pool = av_buffer_pool_init(buf_size, NULL);
>>> -        if (!s->tmp_pool)
>>> +    if (buf_size > s->hevc_headers_size) {
>>> +        av_freep(&s->hevc_headers);
>>> +        s->hevc_headers_size = 0;
>>> +        s->hevc_headers = av_mallocz(buf_size);
>>>
>>
>> The earlier code did not zero-initialize the buffer. I thought the
>> set_[vsp]ps functions were enough to initialize them.
>>
> 
> Better to be on the safe side, in case I've forgotten to set a reserved
> field (or better yet, I should remove references to them in case they're
> renamed.
> 
> 
>>> +        if (!s->hevc_headers)
>>>  return AVERROR(ENOMEM);
>>> -        s->tmp_pool_ele_size = buf_size;
>>> +        s->hevc_headers_size = buf_size;
>>>  }
>>>  
>>> -    *data_buf = av_buffer_pool_get(s->tmp_pool);
>>> -    if (!(*data_buf))
>>> -        return AVERROR(ENOMEM);
>>> -
>>>  /* Setup pointers */
>>> -    data_ptr = (*data_buf)->data;
>>> -    hdr = (HEVCHeaderSet *)data_ptr;
>>> +    hdr = s->hevc_headers;
>>> +    data_ptr = (uintptr_t)hdr;
>>>  hdr->hvps = (HEVCHeaderVPS *)(data_ptr + base_size);
>>>
>>
>> The standard way to get pointer that is at a certain byte offset from a
>> given pointer is by casting said pointer to char* and applying the
>> offset; casting to uintptr_t and back will work, but is actually not
>> guaranteed. In fact, the whole data_ptr variable seems unnecessary.
>> You may use FF_FIELD_AT from lavu/internal.h if you think that it is
>> helpful for this. (I don't think so.)
>>
> 
> Couldn't really think of a way to remove the data_ptr field.
> Do you have any ideas? If not, I'll push this later today.

Ok, I now see that data_ptr is used for more than one line and in a way
which makes it hard to avoid. So keep it, but make it char*.

- Andreas