<div dir="ltr">Víctor, <div> </div><div>Thanks for the hints. I'll check gstreamer-vaapi code and try to hack it in the way as you suggested.</div></div><div class="gmail_extra"> <div class="gmail_quote">On Wed, May 25, 2016 at 3:27 AM, Víctor M. Jáquez L. <<a href="mailto:vjaquez@igalia.com" target="_blank">vjaquez@igalia.com</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">On 05/25/16 at 09:45am, Víctor M. Jáquez L. wrote: > On 05/25/16 at 01:22am, Anon wrote: > > Hi All, > > > > I want to implement the following use case on Intel Haswell CPU with HD > > graphics: > > > > offscreen OpenGL rendering to OpenGL 2D texture -> use > > EGL_MESA_image_dma_buf_export > > to export the texture as prime fd -> vaCreateSurfaces from the prime fd -> > > use vaapi to hw acclerate h264 encoding > > > > Is this supported by the latest va-api and libva-intel-driver release > > (i.e., 1.7.0)? If so, is there any example I can start with? > > > > I did read the sample application h264encode.c and thought it might be a > > good starting point. However, the first major issue I encountered is when > > creating context through vaCreateContext, a number of pre-allocated va > > surfaces are required and thus statically associated with the new va > > context. However, with prime fd, the fd will change so a new va surface is > > created with every new prime fd. I don't see any API can be used to > > dynamically add/remove va surfaces after the context is created. Can you > > please give me some suggestions? > > You can look at the gstreamer-vaapi[1] code, which supports it, though right > now is under discussion and rewriting. > > 1. <a href="https://cgit.freedesktop.org/gstreamer/gstreamer-vaapi" rel="noreferrer" target="_blank">https://cgit.freedesktop.org/gstreamer/gstreamer-vaapi</a> Ugh... morning coffee is not kicking yet. This specific pipeline is not supported out-of-the-box, perhaps hacking gldownload element to export dmabuf elements from a specific texture grabber by glsrcbin. I don't know the details. Anyway, from the VA side, afaik, encoders/decoders need pre-allocate the output surfaces, not the input ones. You can associate the prime fd to a surface and encode it. Don't release the fd until the surface is already processed. vmjl </blockquote></div> </div>