diff mbox series

[FFmpeg-devel] avcodec/nvenc: Rework on DTS generation

Message ID tencent_42C2E7D12AB5D3109C6ECB97A62CD08C2606@qq.com
State New
Headers show
Series [FFmpeg-devel] avcodec/nvenc: Rework on DTS generation | expand

Checks

Context Check Description
andriy/make_x86 success Make finished
andriy/make_fate_x86 success Make fate finished

Commit Message

Zhao Zhili July 17, 2024, 9:47 a.m. UTC
From: Zhao Zhili <zhilizhao@tencent.com>

Before the patch, the method to generate DTS only works with
timebase equal to 1/fps. With timebase like 1/1000

./ffmpeg -i foo.mp4 -an -c:v h264_nvenc -enc_time_base 1/1000 bar.mp4

pts 0    dts -3
pts 160  dts 37
pts 80   dts 77
pts 40   dts 117 <-- invalid
pts 120  dts 157
pts 320  dts 197
pts 240  dts 237
pts 200  dts 277 <-- invalid
pts 280  dts 317 <-- invalid

The generated DTS can be larger than PTS, since it only reorder the
input PTS and minus the number of frame delay, which doesn't take
timebase into account. It should minus the "time" of frame delay.

9a245bd trying to fix the issue, but the implementation is incomplete,
which only use time_base.num. Then it got reverted by ac7c265b33b.

After this patch:

pts 0    dts -120
pts 160  dts -80
pts 80   dts -40
pts 40   dts 0
pts 120  dts 40
pts 320  dts 80
pts 240  dts 120
pts 200  dts 160
pts 280  dts 200
---
 libavcodec/nvenc.c | 62 ++++++++++++++++++++++++++++++++++++++--------
 libavcodec/nvenc.h |  3 +++
 2 files changed, 55 insertions(+), 10 deletions(-)

Comments

Timo Rothenpieler July 17, 2024, 10:03 a.m. UTC | #1
On 17/07/2024 11:47, Zhao Zhili wrote:
> From: Zhao Zhili <zhilizhao@tencent.com>
> 
> Before the patch, the method to generate DTS only works with
> timebase equal to 1/fps. With timebase like 1/1000
> 
> ./ffmpeg -i foo.mp4 -an -c:v h264_nvenc -enc_time_base 1/1000 bar.mp4
> 
> pts 0    dts -3
> pts 160  dts 37
> pts 80   dts 77
> pts 40   dts 117 <-- invalid
> pts 120  dts 157
> pts 320  dts 197
> pts 240  dts 237
> pts 200  dts 277 <-- invalid
> pts 280  dts 317 <-- invalid
> 
> The generated DTS can be larger than PTS, since it only reorder the
> input PTS and minus the number of frame delay, which doesn't take
> timebase into account. It should minus the "time" of frame delay.
> 
> 9a245bd trying to fix the issue, but the implementation is incomplete,
> which only use time_base.num. Then it got reverted by ac7c265b33b.
> 
> After this patch:
> 
> pts 0    dts -120
> pts 160  dts -80
> pts 80   dts -40
> pts 40   dts 0
> pts 120  dts 40
> pts 320  dts 80
> pts 240  dts 120
> pts 200  dts 160
> pts 280  dts 200
> ---
>   libavcodec/nvenc.c | 62 ++++++++++++++++++++++++++++++++++++++--------
>   libavcodec/nvenc.h |  3 +++
>   2 files changed, 55 insertions(+), 10 deletions(-)
> 
> diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
> index ab92395ed6..34448462f0 100644
> --- a/libavcodec/nvenc.c
> +++ b/libavcodec/nvenc.c
> @@ -1893,7 +1893,8 @@ static av_cold int nvenc_setup_surfaces(AVCodecContext *avctx)
>       if (!ctx->frame_data_array)
>           return AVERROR(ENOMEM);
>   
> -    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces, sizeof(int64_t), 0);
> +    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces + ctx->encode_config.frameIntervalP,
> +                                         sizeof(int64_t), 0);
>       if (!ctx->timestamp_list)
>           return AVERROR(ENOMEM);
>   
> @@ -2347,26 +2348,65 @@ static inline int64_t timestamp_queue_dequeue(AVFifo *queue)
>       return timestamp;
>   }
>   
> +static inline int64_t timestamp_queue_peek(AVFifo *queue, size_t index)
> +{
> +    int64_t timestamp = AV_NOPTS_VALUE;
> +    av_fifo_peek(queue, &timestamp, 1, index);
> +
> +    return timestamp;
> +}
> +
>   static int nvenc_set_timestamp(AVCodecContext *avctx,
>                                  NV_ENC_LOCK_BITSTREAM *params,
>                                  AVPacket *pkt)
>   {
>       NvencContext *ctx = avctx->priv_data;
> +    int delay;
> +    int64_t delay_time;
>   
>       pkt->pts = params->outputTimeStamp;
>   
> -    if (avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER) {
> -FF_DISABLE_DEPRECATION_WARNINGS
> -        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list) -
> -#if FF_API_TICKS_PER_FRAME
> -            FFMAX(avctx->ticks_per_frame, 1) *
> -#endif
> -            FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
> -FF_ENABLE_DEPRECATION_WARNINGS
> -    } else {
> +    if (!(avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER)) {
>           pkt->dts = pkt->pts;
> +        return 0;
> +    }
> +
> +    // This can be more than necessary, but we don't know the real reorder delay.
> +    delay = FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
> +    if (ctx->output_frame_num >= delay) {
> +        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list);
> +        ctx->output_frame_num++;
> +        return 0;
>       }
>   
> +    delay_time = ctx->initial_delay_time;
> +    if (!delay_time) {
> +        int64_t t1, t2, t3;
> +        t1 = timestamp_queue_peek(ctx->timestamp_list, delay);
> +        t2 = timestamp_queue_peek(ctx->timestamp_list, 0);
> +        t3 = (delay > 1) ? timestamp_queue_peek(ctx->timestamp_list, 1) : t1;
> +
> +        if (t1 != AV_NOPTS_VALUE) {
> +            delay_time = t1 - t2;
> +        } else if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
> +            delay_time = av_rescale_q(delay, (AVRational) {avctx->framerate.den, avctx->framerate.num},
> +                                      avctx->time_base);
> +        } else if (t3 != AV_NOPTS_VALUE) {
> +            delay_time = delay * (t3 - t2);
> +        } else {
> +            delay_time = delay;
> +        }
> +        ctx->initial_delay_time = delay_time;
> +    }
> +
> +    /* The following method is simple, but doesn't guarantee monotonic with VFR
> +     * when delay_time isn't accurate (that is, t1 == AV_NOPTS_VALUE)
> +     *
> +     * dts = timestamp_queue_peek(ctx->timestamp_list, ctx->output_frame_num) - delay_time
> +     */
> +    pkt->dts = timestamp_queue_peek(ctx->timestamp_list, 0) - delay_time * (delay - ctx->output_frame_num) / delay;
> +    ctx->output_frame_num++;
> +
>       return 0;
>   }
>   
> @@ -2902,4 +2942,6 @@ av_cold void ff_nvenc_encode_flush(AVCodecContext *avctx)
>   
>       nvenc_send_frame(avctx, NULL);
>       av_fifo_reset2(ctx->timestamp_list);
> +    ctx->output_frame_num = 0;
> +    ctx->initial_delay_time = 0;
>   }
> diff --git a/libavcodec/nvenc.h b/libavcodec/nvenc.h
> index 09de00badc..dc7fe41951 100644
> --- a/libavcodec/nvenc.h
> +++ b/libavcodec/nvenc.h
> @@ -206,6 +206,9 @@ typedef struct NvencContext
>       AVFifo *output_surface_queue;
>       AVFifo *output_surface_ready_queue;
>       AVFifo *timestamp_list;
> +    // This is for DTS calculating, reset after flush
> +    int64_t output_frame_num;
> +    int64_t initial_delay_time;
>   
>       NV_ENC_SEI_PAYLOAD *sei_data;
>       int sei_data_size;

This looks good on first glance, but I'm away from any nvenc capable PC 
for the next two weeks, so it'll unfortunately be a while until I can 
properly test and apply.


Timo
Timo Rothenpieler Aug. 2, 2024, 4:28 p.m. UTC | #2
applied, thanks
Dennis Mungai Aug. 2, 2024, 4:54 p.m. UTC | #3
On Fri, 2 Aug 2024, 19:28 Timo Rothenpieler, <timo@rothenpieler.org> wrote:

> applied, thanks
>


This is a bugfix, can it be backported to release 6.x and 7.x?

>
Timo Rothenpieler Aug. 2, 2024, 5:38 p.m. UTC | #4
On 02.08.2024 18:54, Dennis Mungai wrote:
> On Fri, 2 Aug 2024, 19:28 Timo Rothenpieler, <timo@rothenpieler.org> wrote:
> 
>> applied, thanks
>>
> 
> 
> This is a bugfix, can it be backported to release 6.x and 7.x?

It's also a quite substantial change in behaviour, so I'd like to wait 
with that for a while.
Andreas Rheinhardt Aug. 7, 2024, 10:18 a.m. UTC | #5
Zhao Zhili:
> From: Zhao Zhili <zhilizhao@tencent.com>
> 
> Before the patch, the method to generate DTS only works with
> timebase equal to 1/fps. With timebase like 1/1000
> 
> ./ffmpeg -i foo.mp4 -an -c:v h264_nvenc -enc_time_base 1/1000 bar.mp4
> 
> pts 0    dts -3
> pts 160  dts 37
> pts 80   dts 77
> pts 40   dts 117 <-- invalid
> pts 120  dts 157
> pts 320  dts 197
> pts 240  dts 237
> pts 200  dts 277 <-- invalid
> pts 280  dts 317 <-- invalid
> 
> The generated DTS can be larger than PTS, since it only reorder the
> input PTS and minus the number of frame delay, which doesn't take
> timebase into account. It should minus the "time" of frame delay.
> 
> 9a245bd trying to fix the issue, but the implementation is incomplete,
> which only use time_base.num. Then it got reverted by ac7c265b33b.
> 
> After this patch:
> 
> pts 0    dts -120
> pts 160  dts -80
> pts 80   dts -40
> pts 40   dts 0
> pts 120  dts 40
> pts 320  dts 80
> pts 240  dts 120
> pts 200  dts 160
> pts 280  dts 200

It looks like you only have two reorder frames here, so shouldn't the
first dts be -80?

> ---
>  libavcodec/nvenc.c | 62 ++++++++++++++++++++++++++++++++++++++--------
>  libavcodec/nvenc.h |  3 +++
>  2 files changed, 55 insertions(+), 10 deletions(-)
> 
> diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
> index ab92395ed6..34448462f0 100644
> --- a/libavcodec/nvenc.c
> +++ b/libavcodec/nvenc.c
> @@ -1893,7 +1893,8 @@ static av_cold int nvenc_setup_surfaces(AVCodecContext *avctx)
>      if (!ctx->frame_data_array)
>          return AVERROR(ENOMEM);
>  
> -    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces, sizeof(int64_t), 0);
> +    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces + ctx->encode_config.frameIntervalP,
> +                                         sizeof(int64_t), 0);
>      if (!ctx->timestamp_list)
>          return AVERROR(ENOMEM);
>  
> @@ -2347,26 +2348,65 @@ static inline int64_t timestamp_queue_dequeue(AVFifo *queue)
>      return timestamp;
>  }
>  
> +static inline int64_t timestamp_queue_peek(AVFifo *queue, size_t index)
> +{
> +    int64_t timestamp = AV_NOPTS_VALUE;
> +    av_fifo_peek(queue, &timestamp, 1, index);
> +
> +    return timestamp;
> +}
> +
>  static int nvenc_set_timestamp(AVCodecContext *avctx,
>                                 NV_ENC_LOCK_BITSTREAM *params,
>                                 AVPacket *pkt)
>  {
>      NvencContext *ctx = avctx->priv_data;
> +    int delay;
> +    int64_t delay_time;
>  
>      pkt->pts = params->outputTimeStamp;
>  
> -    if (avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER) {
> -FF_DISABLE_DEPRECATION_WARNINGS
> -        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list) -
> -#if FF_API_TICKS_PER_FRAME
> -            FFMAX(avctx->ticks_per_frame, 1) *
> -#endif
> -            FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
> -FF_ENABLE_DEPRECATION_WARNINGS
> -    } else {
> +    if (!(avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER)) {
>          pkt->dts = pkt->pts;
> +        return 0;
> +    }
> +
> +    // This can be more than necessary, but we don't know the real reorder delay.
> +    delay = FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
> +    if (ctx->output_frame_num >= delay) {
> +        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list);
> +        ctx->output_frame_num++;
> +        return 0;
>      }
>  
> +    delay_time = ctx->initial_delay_time;
> +    if (!delay_time) {
> +        int64_t t1, t2, t3;
> +        t1 = timestamp_queue_peek(ctx->timestamp_list, delay);
> +        t2 = timestamp_queue_peek(ctx->timestamp_list, 0);
> +        t3 = (delay > 1) ? timestamp_queue_peek(ctx->timestamp_list, 1) : t1;
> +
> +        if (t1 != AV_NOPTS_VALUE) {
> +            delay_time = t1 - t2;
> +        } else if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
> +            delay_time = av_rescale_q(delay, (AVRational) {avctx->framerate.den, avctx->framerate.num},
> +                                      avctx->time_base);
> +        } else if (t3 != AV_NOPTS_VALUE) {
> +            delay_time = delay * (t3 - t2);
> +        } else {
> +            delay_time = delay;
> +        }
> +        ctx->initial_delay_time = delay_time;
> +    }
> +
> +    /* The following method is simple, but doesn't guarantee monotonic with VFR
> +     * when delay_time isn't accurate (that is, t1 == AV_NOPTS_VALUE)
> +     *
> +     * dts = timestamp_queue_peek(ctx->timestamp_list, ctx->output_frame_num) - delay_time
> +     */
> +    pkt->dts = timestamp_queue_peek(ctx->timestamp_list, 0) - delay_time * (delay - ctx->output_frame_num) / delay;
> +    ctx->output_frame_num++;
> +
>      return 0;
>  }
>  
> @@ -2902,4 +2942,6 @@ av_cold void ff_nvenc_encode_flush(AVCodecContext *avctx)
>  
>      nvenc_send_frame(avctx, NULL);
>      av_fifo_reset2(ctx->timestamp_list);
> +    ctx->output_frame_num = 0;
> +    ctx->initial_delay_time = 0;
>  }
> diff --git a/libavcodec/nvenc.h b/libavcodec/nvenc.h
> index 09de00badc..dc7fe41951 100644
> --- a/libavcodec/nvenc.h
> +++ b/libavcodec/nvenc.h
> @@ -206,6 +206,9 @@ typedef struct NvencContext
>      AVFifo *output_surface_queue;
>      AVFifo *output_surface_ready_queue;
>      AVFifo *timestamp_list;
> +    // This is for DTS calculating, reset after flush
> +    int64_t output_frame_num;
> +    int64_t initial_delay_time;
>  
>      NV_ENC_SEI_PAYLOAD *sei_data;
>      int sei_data_size;
Zhao Zhili Aug. 7, 2024, 11:13 a.m. UTC | #6
> On Aug 7, 2024, at 18:18, Andreas Rheinhardt <andreas.rheinhardt@outlook.com> wrote:
> 
> Zhao Zhili:
>> From: Zhao Zhili <zhilizhao@tencent.com>
>> 
>> Before the patch, the method to generate DTS only works with
>> timebase equal to 1/fps. With timebase like 1/1000
>> 
>> ./ffmpeg -i foo.mp4 -an -c:v h264_nvenc -enc_time_base 1/1000 bar.mp4
>> 
>> pts 0    dts -3
>> pts 160  dts 37
>> pts 80   dts 77
>> pts 40   dts 117 <-- invalid
>> pts 120  dts 157
>> pts 320  dts 197
>> pts 240  dts 237
>> pts 200  dts 277 <-- invalid
>> pts 280  dts 317 <-- invalid
>> 
>> The generated DTS can be larger than PTS, since it only reorder the
>> input PTS and minus the number of frame delay, which doesn't take
>> timebase into account. It should minus the "time" of frame delay.
>> 
>> 9a245bd trying to fix the issue, but the implementation is incomplete,
>> which only use time_base.num. Then it got reverted by ac7c265b33b.
>> 
>> After this patch:
>> 
>> pts 0    dts -120
>> pts 160  dts -80
>> pts 80   dts -40
>> pts 40   dts 0
>> pts 120  dts 40
>> pts 320  dts 80
>> pts 240  dts 120
>> pts 200  dts 160
>> pts 280  dts 200
> 
> It looks like you only have two reorder frames here, so shouldn't the
> first dts be -80?

But NVENC doesn’t provide the reorder delay value. I have added a comment without
changing the method to guess a maximum reorder delay. Maybe we can do better by
parse  the maximum reorder delay from VPS/SPS/PPS, it’s not the patch supposed
to do. I do hope Nvidia to fix their SDK issues than workaround in FFmpeg.

>> +
>> +    // This can be more than necessary, but we don't know the real reorder delay.
>> +    delay = FFMAX(ctx->encode_config.frameIntervalP - 1, 0);

 

> 
>> ---
>> libavcodec/nvenc.c | 62 ++++++++++++++++++++++++++++++++++++++--------
>> libavcodec/nvenc.h |  3 +++
>> 2 files changed, 55 insertions(+), 10 deletions(-)
>> 
>> diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
>> index ab92395ed6..34448462f0 100644
>> --- a/libavcodec/nvenc.c
>> +++ b/libavcodec/nvenc.c
>> @@ -1893,7 +1893,8 @@ static av_cold int nvenc_setup_surfaces(AVCodecContext *avctx)
>>     if (!ctx->frame_data_array)
>>         return AVERROR(ENOMEM);
>> 
>> -    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces, sizeof(int64_t), 0);
>> +    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces + ctx->encode_config.frameIntervalP,
>> +                                         sizeof(int64_t), 0);
>>     if (!ctx->timestamp_list)
>>         return AVERROR(ENOMEM);
>> 
>> @@ -2347,26 +2348,65 @@ static inline int64_t timestamp_queue_dequeue(AVFifo *queue)
>>     return timestamp;
>> }
>> 
>> +static inline int64_t timestamp_queue_peek(AVFifo *queue, size_t index)
>> +{
>> +    int64_t timestamp = AV_NOPTS_VALUE;
>> +    av_fifo_peek(queue, &timestamp, 1, index);
>> +
>> +    return timestamp;
>> +}
>> +
>> static int nvenc_set_timestamp(AVCodecContext *avctx,
>>                                NV_ENC_LOCK_BITSTREAM *params,
>>                                AVPacket *pkt)
>> {
>>     NvencContext *ctx = avctx->priv_data;
>> +    int delay;
>> +    int64_t delay_time;
>> 
>>     pkt->pts = params->outputTimeStamp;
>> 
>> -    if (avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER) {
>> -FF_DISABLE_DEPRECATION_WARNINGS
>> -        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list) -
>> -#if FF_API_TICKS_PER_FRAME
>> -            FFMAX(avctx->ticks_per_frame, 1) *
>> -#endif
>> -            FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
>> -FF_ENABLE_DEPRECATION_WARNINGS
>> -    } else {
>> +    if (!(avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER)) {
>>         pkt->dts = pkt->pts;
>> +        return 0;
>> +    }
>> +
>> +    // This can be more than necessary, but we don't know the real reorder delay.
>> +    delay = FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
>> +    if (ctx->output_frame_num >= delay) {
>> +        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list);
>> +        ctx->output_frame_num++;
>> +        return 0;
>>     }
>> 
>> +    delay_time = ctx->initial_delay_time;
>> +    if (!delay_time) {
>> +        int64_t t1, t2, t3;
>> +        t1 = timestamp_queue_peek(ctx->timestamp_list, delay);
>> +        t2 = timestamp_queue_peek(ctx->timestamp_list, 0);
>> +        t3 = (delay > 1) ? timestamp_queue_peek(ctx->timestamp_list, 1) : t1;
>> +
>> +        if (t1 != AV_NOPTS_VALUE) {
>> +            delay_time = t1 - t2;
>> +        } else if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
>> +            delay_time = av_rescale_q(delay, (AVRational) {avctx->framerate.den, avctx->framerate.num},
>> +                                      avctx->time_base);
>> +        } else if (t3 != AV_NOPTS_VALUE) {
>> +            delay_time = delay * (t3 - t2);
>> +        } else {
>> +            delay_time = delay;
>> +        }
>> +        ctx->initial_delay_time = delay_time;
>> +    }
>> +
>> +    /* The following method is simple, but doesn't guarantee monotonic with VFR
>> +     * when delay_time isn't accurate (that is, t1 == AV_NOPTS_VALUE)
>> +     *
>> +     * dts = timestamp_queue_peek(ctx->timestamp_list, ctx->output_frame_num) - delay_time
>> +     */
>> +    pkt->dts = timestamp_queue_peek(ctx->timestamp_list, 0) - delay_time * (delay - ctx->output_frame_num) / delay;
>> +    ctx->output_frame_num++;
>> +
>>     return 0;
>> }
>> 
>> @@ -2902,4 +2942,6 @@ av_cold void ff_nvenc_encode_flush(AVCodecContext *avctx)
>> 
>>     nvenc_send_frame(avctx, NULL);
>>     av_fifo_reset2(ctx->timestamp_list);
>> +    ctx->output_frame_num = 0;
>> +    ctx->initial_delay_time = 0;
>> }
>> diff --git a/libavcodec/nvenc.h b/libavcodec/nvenc.h
>> index 09de00badc..dc7fe41951 100644
>> --- a/libavcodec/nvenc.h
>> +++ b/libavcodec/nvenc.h
>> @@ -206,6 +206,9 @@ typedef struct NvencContext
>>     AVFifo *output_surface_queue;
>>     AVFifo *output_surface_ready_queue;
>>     AVFifo *timestamp_list;
>> +    // This is for DTS calculating, reset after flush
>> +    int64_t output_frame_num;
>> +    int64_t initial_delay_time;
>> 
>>     NV_ENC_SEI_PAYLOAD *sei_data;
>>     int sei_data_size;
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
diff mbox series

Patch

diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
index ab92395ed6..34448462f0 100644
--- a/libavcodec/nvenc.c
+++ b/libavcodec/nvenc.c
@@ -1893,7 +1893,8 @@  static av_cold int nvenc_setup_surfaces(AVCodecContext *avctx)
     if (!ctx->frame_data_array)
         return AVERROR(ENOMEM);
 
-    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces, sizeof(int64_t), 0);
+    ctx->timestamp_list = av_fifo_alloc2(ctx->nb_surfaces + ctx->encode_config.frameIntervalP,
+                                         sizeof(int64_t), 0);
     if (!ctx->timestamp_list)
         return AVERROR(ENOMEM);
 
@@ -2347,26 +2348,65 @@  static inline int64_t timestamp_queue_dequeue(AVFifo *queue)
     return timestamp;
 }
 
+static inline int64_t timestamp_queue_peek(AVFifo *queue, size_t index)
+{
+    int64_t timestamp = AV_NOPTS_VALUE;
+    av_fifo_peek(queue, &timestamp, 1, index);
+
+    return timestamp;
+}
+
 static int nvenc_set_timestamp(AVCodecContext *avctx,
                                NV_ENC_LOCK_BITSTREAM *params,
                                AVPacket *pkt)
 {
     NvencContext *ctx = avctx->priv_data;
+    int delay;
+    int64_t delay_time;
 
     pkt->pts = params->outputTimeStamp;
 
-    if (avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER) {
-FF_DISABLE_DEPRECATION_WARNINGS
-        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list) -
-#if FF_API_TICKS_PER_FRAME
-            FFMAX(avctx->ticks_per_frame, 1) *
-#endif
-            FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
-FF_ENABLE_DEPRECATION_WARNINGS
-    } else {
+    if (!(avctx->codec_descriptor->props & AV_CODEC_PROP_REORDER)) {
         pkt->dts = pkt->pts;
+        return 0;
+    }
+
+    // This can be more than necessary, but we don't know the real reorder delay.
+    delay = FFMAX(ctx->encode_config.frameIntervalP - 1, 0);
+    if (ctx->output_frame_num >= delay) {
+        pkt->dts = timestamp_queue_dequeue(ctx->timestamp_list);
+        ctx->output_frame_num++;
+        return 0;
     }
 
+    delay_time = ctx->initial_delay_time;
+    if (!delay_time) {
+        int64_t t1, t2, t3;
+        t1 = timestamp_queue_peek(ctx->timestamp_list, delay);
+        t2 = timestamp_queue_peek(ctx->timestamp_list, 0);
+        t3 = (delay > 1) ? timestamp_queue_peek(ctx->timestamp_list, 1) : t1;
+
+        if (t1 != AV_NOPTS_VALUE) {
+            delay_time = t1 - t2;
+        } else if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
+            delay_time = av_rescale_q(delay, (AVRational) {avctx->framerate.den, avctx->framerate.num},
+                                      avctx->time_base);
+        } else if (t3 != AV_NOPTS_VALUE) {
+            delay_time = delay * (t3 - t2);
+        } else {
+            delay_time = delay;
+        }
+        ctx->initial_delay_time = delay_time;
+    }
+
+    /* The following method is simple, but doesn't guarantee monotonic with VFR
+     * when delay_time isn't accurate (that is, t1 == AV_NOPTS_VALUE)
+     *
+     * dts = timestamp_queue_peek(ctx->timestamp_list, ctx->output_frame_num) - delay_time
+     */
+    pkt->dts = timestamp_queue_peek(ctx->timestamp_list, 0) - delay_time * (delay - ctx->output_frame_num) / delay;
+    ctx->output_frame_num++;
+
     return 0;
 }
 
@@ -2902,4 +2942,6 @@  av_cold void ff_nvenc_encode_flush(AVCodecContext *avctx)
 
     nvenc_send_frame(avctx, NULL);
     av_fifo_reset2(ctx->timestamp_list);
+    ctx->output_frame_num = 0;
+    ctx->initial_delay_time = 0;
 }
diff --git a/libavcodec/nvenc.h b/libavcodec/nvenc.h
index 09de00badc..dc7fe41951 100644
--- a/libavcodec/nvenc.h
+++ b/libavcodec/nvenc.h
@@ -206,6 +206,9 @@  typedef struct NvencContext
     AVFifo *output_surface_queue;
     AVFifo *output_surface_ready_queue;
     AVFifo *timestamp_list;
+    // This is for DTS calculating, reset after flush
+    int64_t output_frame_num;
+    int64_t initial_delay_time;
 
     NV_ENC_SEI_PAYLOAD *sei_data;
     int sei_data_size;