[FFmpeg-devel] [PATCH] Vulkan hwcontext and filters

[Mark Thompson]

On 19/01/2020 16:51, Philip Langdale wrote:
> On Sat, 18 Jan 2020 20:23:00 +0000
> Mark Thompson <sw at jkqxz.net> wrote:
> 
>>>  typedef struct AVHWDeviceInternal AVHWDeviceInternal;
>>> diff --git a/libavutil/hwcontext_cuda.c b/libavutil/hwcontext_cuda.c
>>> index 30611b1912..18abb87bbd 100644
>>> --- a/libavutil/hwcontext_cuda.c
>>> +++ b/libavutil/hwcontext_cuda.c
>>> @@ -21,6 +21,9 @@
>>>  #include "hwcontext.h"
>>>  #include "hwcontext_internal.h"
>>>  #include "hwcontext_cuda_internal.h"
>>> +#if CONFIG_VULKAN
>>> +#include "hwcontext_vulkan.h"
>>> +#endif
>>>  #include "cuda_check.h"
>>>  #include "mem.h"
>>>  #include "pixdesc.h"
>>> @@ -42,6 +45,9 @@ static const enum AVPixelFormat
>>> supported_formats[] = { AV_PIX_FMT_YUV444P16,
>>>      AV_PIX_FMT_0RGB32,
>>>      AV_PIX_FMT_0BGR32,
>>> +#if CONFIG_VULKAN
>>> +    AV_PIX_FMT_VULKAN,
>>> +#endif  
>>
>> Do all devices we can do CUDA on also support Vulkan?  If not, this
>> should probably filter it out in get_constraints() to avoid exposing
>> something which can't possibly work.
> 
> I don't have any hardware to test with, but I guess that in theory, we
> do - sufficiently old hardware, with legacy driver branches will
> support cuda and not vulkan, but it's unclear what we could do beyond
> what we already have to handle that. If there's no vulkan support, then
> we will not a vulkan device with a matching UUID and things will error
> out at that stage. At the moment, the only mechanism I can think of to
> detect lack of support is to do the UUID check, but somehow try and do
> it early enough to affect the supported formats, which seems dubious to
> me. So, it'll error out later than you really want, and possibly with a
> slightly unclear message but it'll not do anything wrong.

Fair enough - I agree there isn't much value to adding a lot more code to check it earlier.

Thanks,

- Mark