[Mesa-dev] [PATCH v3 0/9] Gallium Nine

Tue Nov 18 13:21:39 PST 2014

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Am 2014-11-18 19:36, schrieb Henri Verbeet:
> To clarify, that's not what I said. It's mostly just that I'd like
> to see some actual evidence for the (implicit) claim that the
> performance difference is largely due to inherent OpenGL API
> overhead.
I have some microbenchmarks to measure API overhead of d3d and GL here:

https://stefandoesinger.ddns.net/~git/perftest/
(self-signed cert)
Binaries are here:
https://stefandoesinger.ddns.net/~stefan/pts/

I had a chat with mannerov (I guess he's Axel Davy, according to
/whois) on IRC, and he ran the drawprim and clear tests with gallium
nine, wined3d and GL. His results show pretty much the same result in
gallium nine and wined3d, and the GL result handily beats both of them.

Mannerov says his system doesn't show a big difference in real games
between wined3d and nine, so it would be nice to have more test
results from other people who see a big difference in real-world games.

Those benchmarks would be a first step in pointing out overhead
differences between nine and wined3d and give some hints where they
are. The GL result should show if the difference is due to overhead
inside wined3d or the Mesa GL frontend.

Obviously more tests are needed. The tests so far just cover plain
draws, stream source changes, clears and vertex buffer uploads. I
suspect that a major source of overhead are shader changes and / or
shader constant / uniform changes. And so far all of those tests only
cover CPU-side performance problems.