On Thu, Sep 12, 2019 at 4:49 AM Nicolas Dufresne <nicolas@xxxxxxxxxxxx> wrote: > > Le mercredi 11 septembre 2019 à 09:27 +0100, Ezequiel Garcia a écrit : > > On Mon, 2019-09-09 at 16:07 +0900, Tomasz Figa wrote: > > > Hi Ezequiel, > > > > > > On Wed, Sep 4, 2019 at 3:17 AM Ezequiel Garcia <ezequiel@xxxxxxxxxxxxx> wrote: > > > > Hi all, > > > > > > > > This series enables the post-processor support available > > > > on the Hantro G1 VPU. The post-processor block can be > > > > pipelined with the decoder hardware, allowing to perform > > > > operations such as color conversion, scaling, rotation, > > > > cropping, among others. > > > > > > > > The decoder hardware needs its own set of NV12 buffers > > > > (the native decoder format), and the post-processor is the > > > > owner of the CAPTURE buffers. This allows the application > > > > get processed (scaled, converted, etc) buffers, completely > > > > transparently. > > > > > > > > This feature is implemented by exposing other CAPTURE pixel > > > > formats to the application (ENUM_FMT). When the application > > > > sets a pixel format other than NV12, the driver will enable > > > > and use the post-processor transparently. > > > > > > I'll try to review the series a bit later, but a general comment here > > > is that the userspace wouldn't have a way to distinguish between the > > > native and post-processed formats. I'm pretty much sure that > > > post-processing at least imposes some power penalty, so it would be > > > good if the userspace could avoid it if unnecessary. > > > > > > > Hm, that's true, good catch. > > > > So, it would be desirable to retain the current behavior of allowing > > the application to just set a different pixel format and get > > a post-processed frame, transparently. > > > > But at the same time, it would be nice if the application is somehow > > aware of the post-processing happening. Maybe we can expose a more > > accurate media controller topology, have applications enable > > the post-processing pipeline explicitly. > > How it works on the stateful side is that userspace set the encoding > type (the codec), then passes a header (in our case, there will be > parsed structures replacing this) first. The driver then configure > capture format, giving a hint of the "default" or "native" format. This > may or may not be sufficient, but it does work in giving userspace a > hint. The bad side of that is that we can't handle more than 1 native format. For the most backwards-compatible behavior, sorting the results of ENUM_FMT according to format preference would allow the applications that choose the first format returned that works for them to choose the best one. For a further improvement, an ENUM_FMT flag that tells the userspace that a format is preferred could work. That said, modelling the pipeline appropriately using the media controller is the idea I like the most, because it's the most comprehensive solution. That would have to be well specified and documented, though, and sounds like a long term effort. Best regards, Tomasz
