tfogal / NvPipe

NVIDIA-accelerated zero latency video compression library for interactive remoting applications

Introduction

NVPipe is a simple and lightweight library for low-latency video compression. It provides access to NVIDIA's hardware-accelerated codecs as well as a fallback to libx264 (via FFMpeg).

It is a great choice to drastically lower the bandwidth required for your networked interactive client/server application. It would be a poor choice for streaming passive video content.

Sample usage

The library is specifically designed to be easily integratable into existing low-latency streaming applications. NVPipe does not take over any of the network communication aspects, allowing your application to dictate the client/server scenario it is used in.

A sample encoding scenario:

  const size_t width=..., height=...; // input image size.
  const uint8_t* rgba = ...; // your image data.
  size_t osize = ...; // number of bytes in 'output'.
  void* output = malloc(max_osize); // place for the compressed stream.

  // First create our encoder.
  const size_t f_m = 4; // "quality".  Generally in [1,5].
  const size_t fps = 30;
  const uint64_t bitrate = width*height * fps*f_m*0.07;
  nvpipe* enc = nvpipe_create_encoder(NVPIPE_H264_NV, bitrate);

  while(you_have_more_frames) {
    nvpipe_encode(enc, rgba, width*height*4, output,&osize, width, height,
                  NVPIPE_RGBA);

    // Send the frame size and compressed stream to the consuming side.
    uint32_t size = to_network_byte_order_zu(osize);
    send(socket, &size, sizeof(uint32_t), ...);
    send(socket, output, osize, ...);
  }

  // Be sure to destroy the encoder when done.
  nvpipe_destroy(enc);

osize on the encode side will become N in the sample decode scenario:

  nvpipe* dec = nvpipe_create_decoder(NVPIPE_H264_NV);

  size_t width=1920, height=1080;
  size_t imgsz = width * height * 3;
  uint8_t* rgb = malloc(imgsz);

  void* strm = malloc(max_N);

  while(you_have_more_frames) {
    uint32_t N = 0;
    recv(socket, &N, sizeof(uin32_t), ...);
    recv(socket, strm, from_network_byte_order(N), ...);
    nvpipe_decode(dec, strm,N, rgb, width, height);
    use_frame(rgb, width, height); // blit, save; whatever.
  }

  nvpipe_destroy(dec); // destroy the decoder when finished

Note that the underlying H.264 stream does not support alpha channels. Such information would be lost in the scenario above.

Build

This library can optionally make use of the NVIDIA Video Codec SDK headers available from:

https://developer.nvidia.com/nvidia-video-codec-sdk

These headers enable additional features that can improve the quality of the generated stream.

To utilize a custom SDK, set the NVIDIA_VIDEO_CODEC_SDK CMake option to the directory where you have unzipped the SDK. Then build using this library using the standard CMake process.

The only special CMake variable is the boolean

USE_FFMPEG

that controls whether or not the (optional) ffmpeg-based backend is built.

Only shared libraries are supported.

(Optional) FFMpeg-based backend

To use the NVPIPE_H264_NVFFMPEG or NVPIPE_H264_FFMPEG backends, the library must be compiled with FFMpeg support. Furthermore a small modification is required in the FFMpeg source tree: in libavcodec/cuvid.c, change the ulMaxDisplayDelay from 4 to 0.

Then build FFMpeg with the 'nvenc' and 'cuvid' backends enabled, e.g.:

	#!/bin/sh
	export CFLAGS="-ggdb -fPIC"
	export CXXFLAGS="-ggdb -fPIC"
	cu="/usr/local/cuda"
	sdk="${HOME}/sw/nv-video-sdk-7.0.1/"
	ldf="-fPIC"
	ldf="${ldf} -L${cu}/lib64 -Wl,-rpath=${cu}/lib64"
	ldf="${ldf} -L${cu}/lib64/stubs -Wl,-rpath=${cu}/lib64/stubs"
	ldf="${ldf} -L${cu}/lib -Wl,-rpath=${cu}/lib"

	rm -f config.fate config.h config.log config.mak Makefile
	rm -fr libav* libsw*
	../configure \
		--prefix="${HOME}/sw/ffmpeg3.1" \
		--extra-cflags="-I${cu}/include -I${sdk}/include ${CFLAGS}" \
		--extra-cxxflags="-I${cu}/include -I${sdk}/include ${CFLAGS}" \
		--extra-ldflags="${ldf}" \
		--extra-ldexeflags="${ldf}" \
		--extra-ldlibflags="${ldf}" \
		--disable-stripping \
		--assert-level=2 \
		--enable-shared \
		--disable-static \
		--enable-nonfree \
		--enable-nvenc \
		--enable-cuda \
		--enable-cuvid \
		--disable-yasm \
		--disable-gpl \
		--disable-doc \
		--disable-htmlpages \
		--disable-podpages \
		--disable-txtpages \
		--disable-vaapi \
		--disable-everything \
		--enable-decoder=h264 \
		--enable-decoder=h264_cuvid \
		--enable-encoder=nvenc \
		--enable-encoder=nvenc_h264 \
		--enable-encoder=h264_nvenc \
		--enable-encoder=png \
		--enable-hwaccel=h264_cuvid \
		--enable-parser=png \
		--enable-parser=h264 \
		--enable-demuxer=h264 \
		--enable-muxer=mjpeg \
		--enable-muxer=h264 \
		--enable-protocol=pipe \
		--enable-protocol=file \
		--enable-protocol=tcp \
		--enable-protocol=udp \
		--enable-filter=scale \
		--enable-bsf=h264_mp4toannexb
	make -j install

Finally build this library while setting the USE_FFMPEG CMake variable to true.

Supported platforms

NVPipe is supported on both Linux and Windows. Kepler-class NVIDIA hardware is required. The NVIDIA Video Codec SDK is an optional dependency to enable extra features of the generated stream.

OS X support is not plausible in the short term. NVPipe uses the NVIDIA Video Codec SDK under the hood, and the Video Codec SDK is not available on OS X at this time.