On 06/06/18 08:46, Sergio Basurco wrote:> On 05/06/2018 22:23, Mark Thompson wrote: >> On 05/06/18 11:15, Sergio Basurco wrote: >>> I'm decoding h264 with ffmpeg. I want to use the hwaccel decoders. I'm >>> using the cuvid decoder via API. In the fftools code there's a function >>> "hwaccel_retrieve_data" that is supposed to convert the decoded frame >>> (NV12) into any other format, I'm trying YUV420P. >>> >>> The conversion does not return any error, however the resulting data is not >>> correct. Here's the original NV12 frame: >>> >>> https://imgur.com/a/jAb8h12 >>> >>> And here's the conversion to YUV420P (only avframe->data[0] and >>> avframe->data[1] have any data, data[2] is expected to have the V data but >>> it is missing). >>> >>> https://imgur.com/a/ihQeJ0M >>> >>> >>> I think I'm on the right track, based on the code in hwcontext_cuda.c >>> aparently YUV420P is a supported format, I cannot get my head around how I >>> can tell the decoder to convert from NV12 to YUV420P though. >>> >>> Any tips will be appreciated, I'll update if I find anything. >> See the documentation for av_hwframe_transfer_data() >> (<http://ffmpeg.org/doxygen/trunk/hwcontext_8h.html#abf1b1664b8239d953ae2cac8b643815a>), >> in particular: >> >> "If src has an AVHWFramesContext attached, then the format of dst (if set) >> must use one of the formats returned by av_hwframe_transfer_get_formats(src, >> AV_HWFRAME_TRANSFER_DIRECTION_FROM). If dst has an AVHWFramesContext >> attached, then the format of src must use one of the formats returned by >> av_hwframe_transfer_get_formats(dst, AV_HWFRAME_TRANSFER_DIRECTION_TO)" >> >> For CUDA, no conversion during transfer is supported so the only usable >> output format returned by av_hwframe_transfer_get_formats() is the same >> format as the GPU-side frame itself (NV12 in your case): >> <http://git.videolan.org/?p=ffmpeg.git;a=blob;f=libavutil/hwcontext_cuda.c#l176>. > > Hi Mark, > > I did misread that bit of documentation, thanks! > > I'm trying to understand the supported_formats array found in > hwcontext_cuda.c > <https://github.com/FFmpeg/FFmpeg/blob/master/libavutil/hwcontext_cuda.c#L35>. > Can the hw context be initialized so that the decoder returns YUV420P > instead of NV12?
The code there is saying that YUV420P is supported as a format for CUDA frames, but that doesn't mean that it is also supported for decoding to - what formats are supported there is dependent on the decoder hardware. Looking at the decoder code it looks like it only supports NV12 for 8-bit decoding, so YUV420P in CUDA therefore can't be used there (on the other hand, it can be used for encoding and some filter operations). - Mark _______________________________________________ Libav-user mailing list Libav-user@ffmpeg.org http://ffmpeg.org/mailman/listinfo/libav-user
