In your message of 20 January 2000 you write:
> The only significant unresolved issue in the oglbase specification
> is threadsafety (section 3.7). Unfortunately, the traditional SGI way of
> doing this is clearly unacceptable to the Linux kernel guys at this
> time, so the significant overhead of pthread_getspecific() appears
> unavoidable for multithreaded apps.
>
> However, Brian has been doing some experiments with a dispatcher
> that recognizes when the app goes multithreaded and only switches to the
> threadsafe calls at that time, so dispatch overhead for single-threaded
> apps should not be impacted very much.
>
> So the answer seems to be that section 3.7 should require that there
> is a single libGL library which is threadsafe - specifically, that an
> arbitrary (implementation dependent) number of GL contexts may be
> created and may be current to different threads simultaneously.
I'm a little bit opposed to this, as it makes it quite impossible to
get decent performance with a static dispatch table (which our
implementation uses). Wouldn't it be possible to have something
similar to what SUN is using, where you explicitely register that you
want to use thread-safeness. This avoids pretty much every kind of
overhead.
- Thomas
--
Thomas Roell /\ An imperfect plan executed violently
Xi Graphics / \/\ _ is far superior to a perfect plan.
[EMAIL PROTECTED] / / \ \
/ Oelch! \ \ George Patton
