[FFmpeg-devel] [PATCH] lavc/cuvid: fail early if GPU can't handle video parameters

Pavel Koshevoy pkoshevoy at gmail.com
Sun Jan 22 19:09:13 EET 2017


On 01/22/2017 05:28 AM, Mark Thompson wrote:
> On 21/01/17 17:35, Pavel Koshevoy wrote:
>> On Sat, Jan 21, 2017 at 10:27 AM,  <pkoshevoy at gmail.com> wrote:
>>> From: Pavel Koshevoy <pkoshevoy at gmail.com>
>>>
>>> Evidently CUVID doesn't support decoding 422 or 444 chroma formats,
>>> and only a limited set of resolutions per codec are supported.
>>>
>>> Unfortunately CUVID silently drops packets for video of unsupported
>>> resolution.  However, it will error-out at cuvidCreateDecoder call
>>> if the indicated video resolution is not supported.
>>>
>>> Given that stream resolution and pixel format are typically known as
>>> a result of probing it is better to use this information during
>>> avcodec_open2 call to fail immediately, rather than proceeding to
>>> decode and never receiving any frames from the decoder nor receiving
>>> any indication of decode failure.
>>> ---
>>>   libavcodec/cuvid.c | 62 +++++++++++++++++++++++++++++++++++++++++++++++++-----
>>>   1 file changed, 57 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/libavcodec/cuvid.c b/libavcodec/cuvid.c
>>> index 8fc713d..f9c29a1 100644
>>> --- a/libavcodec/cuvid.c
>>> +++ b/libavcodec/cuvid.c
>>> @@ -612,7 +612,11 @@ static av_cold int cuvid_decode_end(AVCodecContext *avctx)
>>>       return 0;
>>>   }
>>>
>>> -static int cuvid_test_dummy_decoder(AVCodecContext *avctx, CUVIDPARSERPARAMS *cuparseinfo)
>>> +static int cuvid_test_dummy_decoder(AVCodecContext *avctx,
>>> +                                    const CUVIDPARSERPARAMS *cuparseinfo,
>>> +                                    cudaVideoChromaFormat probed_chroma_format,
>>> +                                    int probed_width,
>>> +                                    int probed_height)
>>>   {
>>>       CuvidContext *ctx = avctx->priv_data;
>>>       CUVIDDECODECREATEINFO cuinfo;
>>> @@ -622,11 +626,11 @@ static int cuvid_test_dummy_decoder(AVCodecContext *avctx, CUVIDPARSERPARAMS *cu
>>>       memset(&cuinfo, 0, sizeof(cuinfo));
>>>
>>>       cuinfo.CodecType = cuparseinfo->CodecType;
>>> -    cuinfo.ChromaFormat = cudaVideoChromaFormat_420;
>>> +    cuinfo.ChromaFormat = probed_chroma_format;
>>>       cuinfo.OutputFormat = cudaVideoSurfaceFormat_NV12;
>>>
>>> -    cuinfo.ulWidth = 1280;
>>> -    cuinfo.ulHeight = 720;
>>> +    cuinfo.ulWidth = probed_width;
>>> +    cuinfo.ulHeight = probed_height;
>>>       cuinfo.ulTargetWidth = cuinfo.ulWidth;
>>>       cuinfo.ulTargetHeight = cuinfo.ulHeight;
>>>
>>> @@ -653,6 +657,36 @@ static int cuvid_test_dummy_decoder(AVCodecContext *avctx, CUVIDPARSERPARAMS *cu
>>>       return 0;
>>>   }
>>>
>>> +static int convert_to_cuda_video_chroma_format(enum AVPixelFormat pix_fmt,
>>> +                                               cudaVideoChromaFormat *out)
>>> +{
>>> +    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
>>> +    if (!(out && desc &&
>>> +          (desc->nb_components == 1 || desc->nb_components == 3) &&
>>> +          (desc->log2_chroma_w < 2 && desc->log2_chroma_h < 2)))
>>> +    {
>>> +        return AVERROR(EINVAL);
>>> +    }
>>> +
>>> +    if (desc->nb_components == 1)
>>> +    {
>>> +        *out = cudaVideoChromaFormat_Monochrome;
>>> +    }
>>> +    else if (desc->flags == AV_PIX_FMT_FLAG_PLANAR)
>>> +    {
>>> +        *out = ((desc->log2_chroma_w == 0) ? cudaVideoChromaFormat_444 :
>>> +                (desc->log2_chroma_h == 0) ? cudaVideoChromaFormat_422 :
>>> +                cudaVideoChromaFormat_420);
>>> +    }
>>> +    else
>>> +    {
>>> +        return AVERROR(EINVAL);
>>> +    }
>>> +
>>> +    // unfortunately, 420 is the only one that works:
>>> +    return (*out == cudaVideoChromaFormat_420) ? 0 : AVERROR_EXTERNAL;
>>> +}
>>> +
>>>   static av_cold int cuvid_decode_init(AVCodecContext *avctx)
>>>   {
>>>       CuvidContext *ctx = avctx->priv_data;
>>> @@ -663,12 +697,27 @@ static av_cold int cuvid_decode_init(AVCodecContext *avctx)
>>>       CUcontext cuda_ctx = NULL;
>>>       CUcontext dummy;
>>>       const AVBitStreamFilter *bsf;
>>> +    cudaVideoChromaFormat probed_chroma_format;
>>> +    int probed_width;
>>> +    int probed_height;
>>>       int ret = 0;
>>>
>>>       enum AVPixelFormat pix_fmts[3] = { AV_PIX_FMT_CUDA,
>>>                                          AV_PIX_FMT_NV12,
>>>                                          AV_PIX_FMT_NONE };
>>>
>>> +    enum AVPixelFormat probed_pix_fmt = (avctx->pix_fmt <= 0 ?
>>> +                                         AV_PIX_FMT_YUV420P :
>>> +                                         avctx->pix_fmt);
>>> +
>>> +    ret = convert_to_cuda_video_chroma_format(probed_pix_fmt, &probed_chroma_format);
>>> +    if (ret < 0) {
>>> +        // pixel format is not supported:
>>> +        return ret;
>>> +    }
>>> +    probed_width = avctx->coded_width ? avctx->coded_width : 1280;
>>> +    probed_height = avctx->coded_height ? avctx->coded_height : 720;
>>> +
>>>       // Accelerated transcoding scenarios with 'ffmpeg' require that the
>>>       // pix_fmt be set to AV_PIX_FMT_CUDA early. The sw_pix_fmt, and the
>>>       // pix_fmt for non-accelerated transcoding, do not need to be correct
>>> @@ -824,7 +873,10 @@ static av_cold int cuvid_decode_init(AVCodecContext *avctx)
>>>       if (ret < 0)
>>>           goto error;
>>>
>>> -    ret = cuvid_test_dummy_decoder(avctx, &ctx->cuparseinfo);
>>> +    ret = cuvid_test_dummy_decoder(avctx, &ctx->cuparseinfo,
>>> +                                   probed_chroma_format,
>>> +                                   probed_width,
>>> +                                   probed_height);
>>>       if (ret < 0)
>>>           goto error;
>>>
>>> --
>>> 2.9.2
>>>
>>
>> Despite wm4 objections I am resubmitting this patch, because I can
>> find no other reasonable method for rejecting ffmpeg cuvid decoder
>> when it can't handle a given video stream.
> An alternative method would be to follow the approach of the qsv decoder.  There, it doesn't open anything until enough information has been parsed from the stream to have all the appropriate parameters.  That entails making a parser inside the decoder and ensuring that the parser returns all of the information you need to pass to the decoder creation, but it avoids any possibility that the user doesn't supply the necessary information and also allows stream format changes to be handled without external support.
>


That sounds interesting, although it's an overkill for my problem -- I already 
have the probed stream codec parameters, and it is wasteful not to use them to 
initialize the codec.  Your proposed solution sounds robust, but also more 
difficult to implement.  I may attempt it (I'll have to see how QSV does it), 
but I my patch would still be an improvement if it were applied now.

Pavel.


More information about the ffmpeg-devel mailing list