Re: [PATCH 0/2] fbdev: add fillrect and copyarea ioctls

[Dave Airlie <airlied@gmail.com>]

On Tue, Jul 14, 2009 at 1:43 PM, Ali Gholami Rudi<ali@rudi.ir> wrote:
> Dave Airlie <airlied@gmail.com> wrote:
>> On Tue, Jul 14, 2009 at 2:25 AM, Ali Gholami Rudi<ali@rudi.ir> wrote:
>> > Arjan van de Ven <arjan@infradead.org> wrote:
>> >> can we turn this around, is there a reason to add them?
>> >> or in other words, how / where would these be used ?
>> >
>> > User-space programs that use framebuffer directly can use them.  I was
>> > writing a simple framebuffer virtual terminal (using libfreetype for
>> > fonts; like fbterm); scrolling and painting boxes would be faster if
>> > there was someway of using hardware accelerated operations.  I think
>> > other similar programs can benefit, too.
>>
>> The general opinion is we should keep acceleration in userspace if at
>> all possible.
>
> I see.  The line between user- and kernel-space for graphic applications
> is very blurred to me :-)
>
>> Not all hw can implement these usefully in the kernel, directfb
>
> AFAICT, many major ones like intelfb, radeonfb and nv implement them and
> those that can't, use a software implementation.  You mean they are
> unreliable or that there is little performance improvement because of
> the way those operations are implemented in the kernel?

Generally newer consumer HW can't implement a simple blit without using the
3D engine which requires a large amount of state to be sent to the device.
Keeping the state encoding and stuff out of the kernel and in userspace is
generally seen as the right thing to do.

>> already does some
>> things for this.
>
> Sometime ago I did try running directfb on a radeon r300 and it failed.
> I didn't try hard to see what's wrong but I got the impression that
> userspace apps are not good at using the hardware directly.

directfb has hw accel for different cards, really porting the code out of
the kernel for the cards you are interested in would probably benefit
more.

Dave.