[FFmpeg-devel] [PATCH] ffmpeg_qsv.c: Init an hwframes_context for decoder instead of encoder

wm4 nfxjfg at googlemail.com
Fri Jan 20 12:05:33 EET 2017


On Fri, 20 Jan 2017 17:41:01 +0800
"Huang, Zhengxu" <zhengxu.maxwell at gmail.com> wrote:

> From 2149f87637ab941be14828f7ae2c224908784c7d Mon Sep 17 00:00:00 2001
> From: Zhengxu <zhengxu.maxwell at gmail.com>
> Date: Wed, 4 Jan 2017 16:43:43 +0800
> Subject: [PATCH] ffmpeg_qsv.c: Init an hwframes_context for decoder instead of
>  encoder.
> 
> We consider that encoder is the last stage in the pipeline. Thinking
> about filters, they get hwframes_context from their inputs. Likewise,
> encoder should get hwframes_context from its input instead creating a
> new faker one. Encoder can get acuurate parameters by doing so.

the idea is right. What ffmpeg.c does is a gross hack that avconv
doesn't have or need. Ideally, ffmpeg_qsv.c would look like avconv's
equivalent file:

https://git.libav.org/?p=libav.git;a=blob;f=avconv_qsv.c;hb=HEAD

(It's quite possible that this still lacks avconv changes needed for
this, and which were either skipped during merge, or not merged yet.)

> Signed-off-by: ChaoX A Liu <chaox.a.liu at gmail.com>
> Signed-off-by: Huang, Zhengxu <zhengxu.maxwell at gmail.com>
> Signed-off-by: Andrew, Zhang <huazh407 at gmail.com>
> ---
>  ffmpeg_qsv.c | 95 +++++++++++++++++++++++++++---------------------------------
>  1 file changed, 43 insertions(+), 52 deletions(-)
> 
> diff --git a/ffmpeg_qsv.c b/ffmpeg_qsv.c
> index 86824b6..8cedb7e 100644
> --- a/ffmpeg_qsv.c
> +++ b/ffmpeg_qsv.c
> @@ -81,25 +81,26 @@ int qsv_init(AVCodecContext *s)
>              return ret;
>      }
>  
> -    av_buffer_unref(&ist->hw_frames_ctx);
> -    ist->hw_frames_ctx = av_hwframe_ctx_alloc(hw_device_ctx);
> -    if (!ist->hw_frames_ctx)
> -        return AVERROR(ENOMEM);
> -
> -    frames_ctx   = (AVHWFramesContext*)ist->hw_frames_ctx->data;
> -    frames_hwctx = frames_ctx->hwctx;
> -
> -    frames_ctx->width             = FFALIGN(s->coded_width,  32);
> -    frames_ctx->height            = FFALIGN(s->coded_height, 32);
> -    frames_ctx->format            = AV_PIX_FMT_QSV;
> -    frames_ctx->sw_format         = s->sw_pix_fmt;
> -    frames_ctx->initial_pool_size = 64;
> -    frames_hwctx->frame_type      = MFX_MEMTYPE_VIDEO_MEMORY_DECODER_TARGET;
> -
> -    ret = av_hwframe_ctx_init(ist->hw_frames_ctx);
> -    if (ret < 0) {
> -        av_log(NULL, AV_LOG_ERROR, "Error initializing a QSV frame pool\n");
> -        return ret;
> +    if(!ist->hw_frames_ctx) {

Why this check? It's pointless - hw_frames_ctx is unset when get_format
is called.

It also changes indentation, which makes reviewing harder.

> +        ist->hw_frames_ctx = av_hwframe_ctx_alloc(hw_device_ctx);
> +        if (!ist->hw_frames_ctx)
> +            return AVERROR(ENOMEM);
> +
> +        frames_ctx   = (AVHWFramesContext*)ist->hw_frames_ctx->data;
> +        frames_hwctx = frames_ctx->hwctx;
> +
> +        frames_ctx->width             = FFALIGN(s->coded_width,  32);
> +        frames_ctx->height            = FFALIGN(s->coded_height, 32);
> +        frames_ctx->format            = AV_PIX_FMT_QSV;
> +        frames_ctx->sw_format         = s->sw_pix_fmt;
> +        frames_ctx->initial_pool_size = 64;
> +        frames_hwctx->frame_type      = MFX_MEMTYPE_VIDEO_MEMORY_DECODER_TARGET;
> +
> +        ret = av_hwframe_ctx_init(ist->hw_frames_ctx);
> +        if (ret < 0) {
> +            av_log(NULL, AV_LOG_ERROR, "Error initializing a QSV frame pool\n");
> +            return ret;
> +        }
>      }
>  
>      ist->hwaccel_get_buffer = qsv_get_buffer;
> @@ -114,9 +115,8 @@ int qsv_transcode_init(OutputStream *ost)
>      const enum AVPixelFormat *pix_fmt;
>  
>      int err, i;
> -    AVBufferRef *encode_frames_ref = NULL;
> -    AVHWFramesContext *encode_frames;
> -    AVQSVFramesContext *qsv_frames;
> +    AVHWFramesContext *frames_ctx;
> +    AVQSVFramesContext *frames_hwctx;
>  
>      /* check if the encoder supports QSV */
>      if (!ost->enc->pix_fmts)
> @@ -150,42 +150,33 @@ int qsv_transcode_init(OutputStream *ost)
>      if (!hw_device_ctx) {
>          err = qsv_device_init(ist);
>          if (err < 0)
> -            goto fail;
> -    }
> -
> -    // This creates a dummy hw_frames_ctx for the encoder to be
> -    // suitably initialised.  It only contains one real frame, so
> -    // hopefully doesn't waste too much memory.
> -
> -    encode_frames_ref = av_hwframe_ctx_alloc(hw_device_ctx);
> -    if (!encode_frames_ref) {
> -        err = AVERROR(ENOMEM);
> -        goto fail;
> +            return err;
>      }
> -    encode_frames = (AVHWFramesContext*)encode_frames_ref->data;
> -    qsv_frames = encode_frames->hwctx;
>  
> -    encode_frames->width     = FFALIGN(ist->resample_width,  32);
> -    encode_frames->height    = FFALIGN(ist->resample_height, 32);
> -    encode_frames->format    = AV_PIX_FMT_QSV;
> -    encode_frames->sw_format = AV_PIX_FMT_NV12;
> -    encode_frames->initial_pool_size = 1;
> +    ist->dec_ctx->pix_fmt = AV_PIX_FMT_QSV;
> +    ist->resample_pix_fmt = AV_PIX_FMT_QSV;

These lines should be removed. AFAIK they're hacks to make ffmpeg.c
"behave". Please fix the root of the issue instead.

> +    ist->hw_frames_ctx    = av_hwframe_ctx_alloc(hw_device_ctx);
> +    if (!ist->hw_frames_ctx)
> +        return AVERROR(ENOMEM);
>  
> -    qsv_frames->frame_type = MFX_MEMTYPE_VIDEO_MEMORY_DECODER_TARGET;
> +    frames_ctx   = (AVHWFramesContext*)ist->hw_frames_ctx->data;
> +    frames_hwctx = frames_ctx->hwctx;
>  
> -    err = av_hwframe_ctx_init(encode_frames_ref);
> -    if (err < 0)
> -        goto fail;
> +    frames_ctx->width             = FFALIGN(ist->resample_width,  32);
> +    frames_ctx->height            = FFALIGN(ist->resample_height, 32);
> +    frames_ctx->format            = AV_PIX_FMT_QSV;
> +    frames_ctx->sw_format         = AV_PIX_FMT_NV12;
> +    frames_ctx->initial_pool_size = 64;
> +    frames_hwctx->frame_type      = MFX_MEMTYPE_VIDEO_MEMORY_DECODER_TARGET;

Why are you creating another frames_ctx here?

>  
> -    ist->dec_ctx->pix_fmt       = AV_PIX_FMT_QSV;
> -    ist->resample_pix_fmt       = AV_PIX_FMT_QSV;
> +    err = av_hwframe_ctx_init(ist->hw_frames_ctx);
> +    if (err < 0) {
> +        av_buffer_unref(&ist->hw_frames_ctx);
> +        av_log(NULL, AV_LOG_ERROR, "Error initializing a QSV frame pool\n");
> +        return err;
> +    }
>  
> -    ost->enc_ctx->pix_fmt       = AV_PIX_FMT_QSV;
> -    ost->enc_ctx->hw_frames_ctx = encode_frames_ref;
> +    ost->enc_ctx->pix_fmt = AV_PIX_FMT_QSV;

enc_ctx should not be accessed. As you said yourself, filters decide
the encoder input anyway.

>  
>      return 0;
> -
> -fail:
> -    av_buffer_unref(&encode_frames_ref);
> -    return err;
>  }



More information about the ffmpeg-devel mailing list