NVEnc

by rigaya

日本語版はこちら＞＞ 简体中文版本＞＞

This software is meant to investigate performance and image quality of HW encoder (NVENC) of NVIDIA. There are 2 types of software developed, one is command line version that runs independently, and the nother is a output plug-in of Aviutl.

• NVEncC.exe ... Command line version supporting transcoding.
• NVEnc.auo ... Output plugin for Aviutl.

Downloads & update history

rigayaの日記兼メモ帳＞＞
github releases

Install

Install instructions for Windows and Linux

Build

Build instructions for Windows and Linux

System Requirements

Windows 10/11 (x86 / x64)
Linux (x64/aarch64)
Aviutl 1.00 or later (NVEnc.auo)
Hardware which supports NVENC
NVIDIA GPU GeForce Kepler gen or later (GT / GTX 6xx or later)
※ Since GT 63x, 62x etc. are renames of the Fermi generation, they cannot run NVEnc.

Usage and options of NVEncC

Option list and details of NVEncC

Examples of supported encode features

Result of --check-features, a feature list returned from the driver. May depend on the driver version.

Precautions for using NVEnc

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

Main usable functions

Common to NVEnc / NVEncC

• Encoding using NVENC
  • H.264 / AVC
    • YUV420 / YUV444
  • H.265 / HEVC (2nd Gen Maxwell or later)
    • YUV420 / YUV444
    • 10 bits
  • AV1 (Ada Lovelace or later)
• Each encode mode of NVENC
  • CQP (fixed quantization)
  • CBR (Constant bitrate)
  • CBRHQ (Constant bitrate, high quality)
  • VBR (Variable bitrate)
  • VBRHQ (Variable bitrate, high quality)
• Interlaced encoding (by PAFF)
• Lossless output (YUV 420 / YUV 444)
• supports setting of codec profile & level, SAR, colormatrix, maxbitrate, GOP len, etc...

NVEncC

• Supports cuvid decoding
  • MPEG1
  • MPEG2
  • H.264 / AVC
  • HEVC (10 bit / 12bitdepth with YUV444 support)
  • VC-1
  • VP9
  • AV1
• Supports various formats such as avs, vpy, y4m, and raw
• Supports demux/muxing using libavformat
• Supports decode using libavcodec
• Calculation of ssim/psnr/vmaf of the encode
• High performance filtering (VPP, Video Pre-Processing)
  • cuvid built-in hw processing
    • resize
    • deinterlace (normal / bob)
  • GPU filtering by CUDA
    • rff (apply rff flag)
    • deinterlacer
      • afs (Automatic field shift)
      • nnedi
      • yadif
    • decimate
    • mpdecimate
    • colorspace conversion (x64 version only)
      • hdr2sdr
      • lut3d
    • delogo
    • subburn
    • resize
      • bilinear
      • spline16, spline36, spline64
      • lanczos2, lanczos3, lanczos4
      • various algorithms by npp library are available (x64 version only)
      • nvvfx-superres
    • transpose / rotate / flip
    • padding
    • select-every
    • deband
    • noise reduction
      • knn (K-nearest neighbor)
      • pmd (modified pmd method)
      • gauss (npp library, x64 version only)
      • convolution3d
      • nvvfx-artifact-reduction
      • nvvfx-denoise
    • edge / detail enhancement
      • unsharp
      • edgelevel (edge ​​level adjustment)
      • warpsharp

NVEnc.auo (Aviutl plugin)

• Audio encoding
• Mux audio and chapter
• afs (Automatic field shift) support

Auto GPU selection in multi GPU envinronment

NVEncC will automatically select a GPU depending on the options used, when there are multiple GPUs available which support NVENC. --device option can be used to specify on which GPU to run manually.

1. Select GPU which supports...
   Items below will be checked whether the GPU supports it or not

• Codec, Profile, Level
• Additionally, below items will be checked if specified
  • 10bit depth encoding
  • Lossless encoding
  • Interlaced encoding
  • HW decode support for ssim/psnr/vmaf calculation

2. Prefer GPU which supports...

• B frame support

3. If there are multiple GPUs which suports all the items checked in 1. and 2., GPU below will be prefered.

• GPU which has low Video Engine(VE) utilization
• GPU which has low GPU core utilization
• newer Generation GPU
• GPU with more CUDA cores

The purpose of selecting GPU with lower VE/GPU ulitization is to assign tasks to mulitple GPUs and improve the throughput of the tasks. Also, newer Gen GPU and GPU with more cores are assumed to have improved performance.

Please note that VE and GPU ulitization are check at the initialization pahse of the app, and there are delays in values taken. Therefore, it is likely that the multiple tasks started at the same time to run on the same GPU, and divided into multiple GPUs, even if the options are supported in every GPUs.

NVEnc source code

• MIT license.
• This program is based on NVIDA CUDA Samples and includes sample code. This software contains source code provided by NVIDIA Corporation.
• --vpp-nvvfx fiters are power by NVIDIA BROADCAST.
  
• This software depends on jitify, ffmpeg, vmaf, tinyxml2, dtl, libass, ttmath & Caption2Ass. For these licenses, please see the header part of the corresponding source and NVEnc_license.txt.

About source code

Windows ... VC build

Character code: UTF-8-BOM
Line feed: CRLF
Indent: blank x4