[Mesa-dev] [PATCH 09/10] gallium/u_blit: don't release vertex buffer at end of frame / in glFlush

[Jose Fonseca]

----- Original Message -----
> On Wed, Apr 18, 2012 at 5:06 PM, Jose Fonseca <jfonseca at vmware.com>
> wrote:
> > Marek,
> >
> > Sorry for not noticing earlier ( I haven't been able to keep up
> > with mesa traffic lately), but I'm afraid this and the 10/10 are
> > not ok, as on several operating systems, namely, Windows WDDM, it
> > is illegal to flush a command buffer with a buffer mapped.
> 
> The buffer is not mapped when the flush happens. u_blit uses
> pipe_buffer_write_nooverlap, which uses transfer_inline_write, which
> goes into u_default_transfer_inline_write on most (if not all)
> drivers, which unmaps the resource before the function returns.

Ah. OK. Looks good then.

Sorry for jumping on the gun here.
 
> The buffer may be mapped with the 'unsynchronized' flag again later
> though, so drivers should implement that flag to avoid stalls.

Our WDDM winsys has some issues implementing "unsynchronized" semantics, so this may cause some stalls whenever we flush the command buffer, but that is an orthogonal problem we'll need to address independently of this, as this is perfectly valid/common behavior.

Jose