[FFmpeg-devel] [PATCH] avcodec/nvenc: High bit depth encoding for HEVC
Roman Arzumanyan
r.arzumanyan at visionlabs.ai
Thu Apr 18 15:29:30 EEST 2024
Hi Diego,
Asking for my own education.
As far as you've explained, the 8 > 10 bit conversion happens within the
driver, that's understandable.
But how does it influence the output? Does it perform some sort of
proprietary SDR > HDR conversion under the hood that maps the ranges?
What's gonna be the user observable difference between these 2 scenarios?
1) 8 bit input > HEVC 8 bit profile > 8 bit HEVC output
2) 8 bit input > 10 bit up conversion > HEVC 10 bit profile > 10 bit HEVC
output
Better visual quality? Smaller compressed file size?
In other words, what's the purpose of this feature except enabling new
Video Codec SDK capability?
чт, 18 апр. 2024 г. в 13:44, Diego Felix de Souza via ffmpeg-devel <
ffmpeg-devel at ffmpeg.org>:
> Hi Timo,
>
> Thank you for your review. Please check my answers below.
>
> Best regards,
>
> Diego
>
> On 17.04.24, 16:27, "Timo Rothenpieler" <timo at rothenpieler.org> wrote:
>
> External email: Use caution opening links or attachments
>
>
> On 15/04/2024 16:39, Diego Felix de Souza via ffmpeg-devel wrote:
> > From: Diego Felix de Souza <ddesouza at nvidia.com>
> >
> > Adding 10-bit encoding support for HEVC if the input is 8-bit. In
> > case of 8-bit input content, NVENC performs an internal CUDA 8 to
> > 10-bit conversion of the input prior to encoding. Currently, only
> > AV1 supports encoding 8-bit content as 10-bit.
>
> I'm not sure about this one.
> Since it's just "SW", or rather CUDA based, conversion, this job could
> also be done by scale_cuda, outside of some niche formats it doesn't
> support yet.
> Which would also allow for more consistent command lines across versions.
>
>
> Although it is a software-based solution, the conversion is integrated
> with other kernels
> inside the driver. In this way, it will demand fewer memory accesses and
> better Streaming
> Multiprocessor (SM) utilization, leading to a improved performance and a
> lower latency
> compared to the scale_cuda approach. Moreover, the proposed approach
> overall consumes
> less memory since it only requires an 8-bit per channel frame as input and
> no extra 10-bit frames.
>
>
> > Signed-off-by: Diego Felix de Souza <ddesouza at nvidia.com>
> > ---
> > libavcodec/nvenc.c | 8 ++++----
> > libavcodec/nvenc_hevc.c | 1 +
> > 2 files changed, 5 insertions(+), 4 deletions(-)
> >
> > diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
> > index 794174a53f..c302cc7dc4 100644
> > --- a/libavcodec/nvenc.c
> > +++ b/libavcodec/nvenc.c
> > @@ -514,7 +514,7 @@ static int nvenc_check_capabilities(AVCodecContext
> *avctx)
> > }
> >
> > ret = nvenc_check_cap(avctx, NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
> > - if (IS_10BIT(ctx->data_pix_fmt) && ret <= 0) {
> > + if ((IS_10BIT(ctx->data_pix_fmt) || ctx->highbitdepth) && ret <= 0)
> {
> > av_log(avctx, AV_LOG_WARNING, "10 bit encode not supported\n");
> > return AVERROR(ENOSYS);
> > }
> > @@ -1421,7 +1421,7 @@ static av_cold int
> nvenc_setup_hevc_config(AVCodecContext *avctx)
> > }
> >
> > // force setting profile as main10 if input is 10 bit
> > - if (IS_10BIT(ctx->data_pix_fmt)) {
> > + if (IS_10BIT(ctx->data_pix_fmt) || ctx->highbitdepth) {
>
> Won't this need guarded behind the NVENC_HAVE_NEW_BIT_DEPTH_API as well?
> Or would this also work fine with older headers, by just setting this?
>
> For older SDK versions, the HEVC main 10 profile would be selected, but
> the encoded bitstream
> would be still 8-bits. For the sake of consistence and clarity, I will put
> it behind the
> NVENC_HAVE_NEW_BIT_DEPTH_API as you suggested.
>
> > cc->profileGUID = NV_ENC_HEVC_PROFILE_MAIN10_GUID;
> > avctx->profile = AV_PROFILE_HEVC_MAIN_10;
> > }
> > @@ -1435,8 +1435,8 @@ static av_cold int
> nvenc_setup_hevc_config(AVCodecContext *avctx)
> > hevc->chromaFormatIDC = IS_YUV444(ctx->data_pix_fmt) ? 3 : 1;
> >
> > #ifdef NVENC_HAVE_NEW_BIT_DEPTH_API
> > - hevc->inputBitDepth = hevc->outputBitDepth =
> > - IS_10BIT(ctx->data_pix_fmt) ? NV_ENC_BIT_DEPTH_10 :
> NV_ENC_BIT_DEPTH_8;
> > + hevc->inputBitDepth = IS_10BIT(ctx->data_pix_fmt) ?
> NV_ENC_BIT_DEPTH_10 : NV_ENC_BIT_DEPTH_8;
> > + hevc->outputBitDepth = (IS_10BIT(ctx->data_pix_fmt) ||
> ctx->highbitdepth) ? NV_ENC_BIT_DEPTH_10 : NV_ENC_BIT_DEPTH_8;
> > #else
> > hevc->pixelBitDepthMinus8 = IS_10BIT(ctx->data_pix_fmt) ? 2 : 0;
> > #endif
> > diff --git a/libavcodec/nvenc_hevc.c b/libavcodec/nvenc_hevc.c
> > index b949cb1bd7..02e9c9c8eb 100644
> > --- a/libavcodec/nvenc_hevc.c
> > +++ b/libavcodec/nvenc_hevc.c
> > @@ -183,6 +183,7 @@ static const AVOption options[] = {
> > { "fullres", "Two Pass encoding is enabled where first Pass
> is full resolution",
> > 0,
> AV_OPT_TYPE_CONST, { .i64 = NV_ENC_TWO_PASS_FULL_RESOLUTION },
> 0, 0, VE, .unit =
> "multipass" },
> > #endif
> > + { "highbitdepth", "Enable 10 bit encode for 8 bit
> input",OFFSET(highbitdepth),AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
>
> Same question as above, does this always work, even without the new bit
> depth API?
> If not, it also needs the #ifdef.
>
> Same as above, it would give the wrong impression to the user. I will put
> it behind the NVENC_HAVE_NEW_BIT_DEPTH_API.
>
> > #ifdef NVENC_HAVE_LDKFS
> > { "ldkfs", "Low delay key frame scale; Specifies the Scene
> Change frame size increase allowed in case of single frame VBV and CBR",
> >
> OFFSET(ldkfs), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, UCHAR_MAX, VE },
> > --
> > 2.34.1
> >
> >
> -----------------------------------------------------------------------------------
> > NVIDIA GmbH
> > Wuerselen
> > Amtsgericht Aachen
> > HRB 8361
> > Managing Directors: Rebecca Peters, Donald Robertson, Janet Hall, Ludwig
> von Reiche
> >
> >
> -----------------------------------------------------------------------------------
> > This email message is for the sole use of the intended recipient(s) and
> may contain
> > confidential information. Any unauthorized review, use, disclosure or
> distribution
> > is prohibited. If you are not the intended recipient, please contact
> the sender by
> > reply email and destroy all copies of the original message.
> >
> -----------------------------------------------------------------------------------
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> >
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fffmpeg.org%2Fmailman%2Flistinfo%2Fffmpeg-devel&data=05%7C02%7Cddesouza%40nvidia.com%7Cd750b65e093d40e8dd5508dc5eea7d53%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C638489608435687923%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C0%7C%7C%7C&sdata=2F9CxnKyADs6zW1faiZhXQa1FzfcVRKxMJQ8ukoN0nY%3D&reserved=0
> <https://ffmpeg.org/mailman/listinfo/ffmpeg-devel>
> >
> > To unsubscribe, visit link above, or email
> > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>
>
> -----------------------------------------------------------------------------------
> NVIDIA GmbH
> Wuerselen
> Amtsgericht Aachen
> HRB 8361
> Managing Directors: Rebecca Peters, Donald Robertson, Janet Hall, Ludwig
> von Reiche
>
>
> -----------------------------------------------------------------------------------
> This email message is for the sole use of the intended recipient(s) and
> may contain
> confidential information. Any unauthorized review, use, disclosure or
> distribution
> is prohibited. If you are not the intended recipient, please contact the
> sender by
> reply email and destroy all copies of the original message.
>
> -----------------------------------------------------------------------------------
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>
More information about the ffmpeg-devel
mailing list