Skip to content

Hardware Accelerated Decoding

Jump to bottom
WangBin edited this page Mar 16, 2016 · 19 revisions

User

VA-API

Install vainfo libva-dev, libva-intel-vaapi-driver(intel gpu), xvba-va-driver(ATI), vdpau-va-driver(NVIDIA)

If vainfo works, then QtAV should work with VA-API

EGL

Since Qt5.5, you can enable EGL by setting environment var QT_XCB_GL_INTEGRATION=xcb_egl. player and QMLPlayer have EGL configuration too. If EGL is used, VA-API decoding and OpenGL rendering can get the best performance. NVIDIA GPUs currently do not support EGL+VA-API. AMD is not tested.

If EGL is used, you can use zero copy for X11, DRM display. Otherwise, zero copy is only available for X11 and GLX display.

CUDA

OpenGL desktop mode supports 0-copy. To use 0-copy, make sure your program runs with NVIDIA GPU (on Optimus). DirectCopy is the default. DirectCopy mode is fast enough, faster than potplayer, lavfilters+mpc, and may be the fastest one. ZeroCopy mode is even faster.

Linux

nvidia-331 worked. I don't know how to enable CUDA for other drivers. Tell me if you know.

Windows

Install latest nvidia driver.

DXVA

Run in OpenGL ES2 mode (ANGLE) to enable 0-copy. 0-copy in OpenGL desktop mode only supports some new intel GPUs in my tests, maybe also supports AMD.

Developer

Build

Code

CUDA:

QVaraiantHash cuda_opt;
cuda_opt["surfaces"] = 0; //key is property name, case sensitive
cuda_opt["copyMode"] = "ZeroCopy"; // default is "DirectCopy"
QVariantHash opt;
opt["CUDA"] = cuda_opt; //key is decoder name, case sensitive
player.setOptionsForVideoCodec(opt);

Since QtAV1.8, CUDA decoder supports ZeroCopy. Run your program with NVIDIA GPU if you have multiple GPUs.

VA-API:

QVaraiantHash va_opt;
va_opt["display"] = "X11"; //"GLX", "X11", "DRM"
va_opt["copyMode"] = "ZeroCopy"; // "ZeroCopy", "OptimizedCopy", "GenericCopy". Default is "ZeroCopy" if possible
QVariantHash opt;
opt["VAAPI"] = va_opt; //key is decoder name, case sensitive
player.setOptionsForVideoCodec(opt);

Copy Mode

Copy mode for GPU decoders was introduced since 1.5.0. Copy mode can be ZeroCopy, LazyCopy, OptimizedCopy and GenericCopy.

  • ZeroCopy: No data read back from GPU. Directly render the decoded data on GPU. It's the most efficient way. Now it's supported by
    • CUDA
    • VA-API with libva-x11 and libva-glx. It may not work for vdpau backend.
    • VDA(OSX), VideoToolbox(OSX, iOS>=8)
    • DXVA with ANGLE. (Desktop OpenGL is not stable now)
  • LazyCopy: Used in VDA and VideoToolBox
  • OptimizedCopy: Use SSE4.1 for x86 used by DXVA and VA-API(VDA crashes), and NEON for CedarV. SSE4.1 does not work for some GPUs, e.g. NVIDIA GS8400
  • GenericCopy: Use memcpy

Status

  • D3D11
  • Copy
  • EGL (D3D11)
  • DXVA
  • Copy
  • EGL (D3D9)
  • OpenGL (WIP)
  • CUDA
  • Copy
  • EGL (WIP)
  • OpenGL
  • VA-API
  • Copy
  • EGL DMA/DRM
  • EGL TFP
  • GLX TFP
  • VDA
  • Copy
  • OpenGL
  • VideoToolbox
  • OSX Copy/0-Copy
  • iOS Copy/0-Copy
  • VDPAU
  • OMX
Clone this wiki locally