Re: [AD] Poor Ming performance

["Eric Botcazou" <ebotcazou@xxxxxxxxxx>]

> Perhaps the test program should be modified to run the hline() test once
> as it normally does, and once more with the video surface locked during
> the test. What do you guys think?

Then we should also modify all the tests that way ! And the
software-rendered primitives will outperform yet a little more the
hw-accelerated ones.

That's not really the problem: the current hline test gauges the speed of
the single hline() primitive, not of blocks of hline() calls. The overhead
of locking/unlocking must IMHO be taken into account.

But for the filling primitives this overhead is naturally diluted, because
they are actually blocks of hlines() calls and hw-accelerated hlines are
very inefficient in such a situation; hence my proposal.

Btw, I can't figure out why the hw-accelerated hlines are 12 times slower
than the software ones on your system, given the overhead of
locking/unlocking for *each* line.
Are the results the same for vlines and filled rectangles ?
Is anyone else experiencing the same thing ?

--
Eric Botcazou
ebotcazou@xxxxxxxxxx