[FFmpeg-devel] [PATCH v2] avcodec/nvenc: fix flushing for encoder-reuse

[Timo Rothenpieler]

On 16.07.2021 18:09, Jai Luthra wrote:
> Recent encode API restructure (827d6fe73d) removed some state - which broke
> the API for flushing without closing the encoder.
> 
> This functionality was originally added in 3ea7057677 and is useful for
> segmented video, where we don't want to do expensive re-init of HW sessions
> for every segment.
> ---
> v2: Forgot to rebase off master in v1
> 
>   libavcodec/nvenc.c | 16 +++++++++++-----
>   libavcodec/nvenc.h |  2 ++
>   2 files changed, 13 insertions(+), 5 deletions(-)
> 
> diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
> index ee046b9cef..ee0beb4795 100644
> --- a/libavcodec/nvenc.c
> +++ b/libavcodec/nvenc.c
> @@ -2281,6 +2281,11 @@ static int nvenc_send_frame(AVCodecContext *avctx, const AVFrame *frame)
>           return AVERROR(EINVAL);
>   
>       if (frame && frame->buf[0]) {
> +        if (ctx->encoder_flushing) {
> +            ctx->encoder_flushing = 0;
> +            av_fifo_reset(ctx->timestamp_list);
> +        }
> +
>           in_surf = get_free_frame(ctx);
>           if (!in_surf)
>               return AVERROR(EAGAIN);
> @@ -2335,6 +2340,7 @@ static int nvenc_send_frame(AVCodecContext *avctx, const AVFrame *frame)
>           nvenc_codec_specific_pic_params(avctx, &pic_params, ctx->sei_data, sei_count);
>       } else {
>           pic_params.encodePicFlags = NV_ENC_PIC_FLAG_EOS;
> +        ctx->encoder_flushing = 1;
>       }
>   
>       res = nvenc_push_context(avctx);
> @@ -2384,8 +2390,11 @@ int ff_nvenc_receive_packet(AVCodecContext *avctx, AVPacket *pkt)
>   
>       if (!frame->buf[0]) {
>           res = ff_encode_get_frame(avctx, frame);
> -        if (res < 0 && res != AVERROR_EOF)
> +        if (res == AVERROR_EOF || (ctx->encoder_flushing && res == AVERROR(EAGAIN))) {
> +            // flushing mode, continue to send packets
> +        } else if (res < 0) {
>               return res;
> +        }
>       }
>   
>       res = nvenc_send_frame(avctx, frame);
> @@ -2395,7 +2404,7 @@ int ff_nvenc_receive_packet(AVCodecContext *avctx, AVPacket *pkt)
>       } else
>           av_frame_unref(frame);
>   
> -    if (output_ready(avctx, avctx->internal->draining)) {
> +    if (output_ready(avctx, ctx->encoder_flushing)) {
>           av_fifo_generic_read(ctx->output_surface_ready_queue, &tmp_out_surf, sizeof(tmp_out_surf), NULL);
>   
>           res = nvenc_push_context(avctx);
> @@ -2423,8 +2432,5 @@ int ff_nvenc_receive_packet(AVCodecContext *avctx, AVPacket *pkt)
>   
>   av_cold void ff_nvenc_encode_flush(AVCodecContext *avctx)
>   {
> -    NvencContext *ctx = avctx->priv_data;
> -
>       nvenc_send_frame(avctx, NULL);
> -    av_fifo_reset(ctx->timestamp_list);
>   }

Is it really correct for this to do absolutely nothing now?

> diff --git a/libavcodec/nvenc.h b/libavcodec/nvenc.h
> index 85d3a33601..6af3aaf0ce 100644
> --- a/libavcodec/nvenc.h
> +++ b/libavcodec/nvenc.h
> @@ -169,6 +169,8 @@ typedef struct NvencContext
>       NV_ENC_SEI_PAYLOAD *sei_data;
>       int sei_data_size;
>   
> +    int encoder_flushing;
> +
>       struct {
>           void *ptr;
>           int ptr_index;
> 

Can you explain the logic this follows some more? I'm quite confused 
what exactly changed, and how it broke this.
How do you trigger a flush, if calling flush does nothing?

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4494 bytes
Desc: S/MIME Cryptographic Signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20210716/2312e386/attachment.bin>