[Piglit] [PATCH piglit] Time the performance difference with GLX_ARB_context_flush_control

[Eric Anholt]

Neil Roberts <neil at linux.intel.com> writes:

> This adds a test with a heuristic to check whether disabling the
> context flush on release is working. The hope is that a regression
> would give a good indication that a flush has snuck back in.
>
> The test creates a pair of contexts each with its own window and
> renders to both of them with interleaved rendering by switching
> contexts with glXMakeCurrent within each frame. This is done twice,
> once with the flush control of both of the contexts set to flush and
> once with it set to none. The number of frames that can be rendered in
> one second is checked and then compared between the two versions. The
> assumption is that the version which doesn't flush when the context is
> released should be noticeably faster. The test fails if the FPS isn't
> more than 50% faster for the non-flushing version.

If we can do a test without using timing, I'd much prefer that -- timing
tests are really unreliable, in my experience of piglit tests doing so
(and I'm responsible for a couple of them).

I think you could fairly accurately detect an extra flush by looking at
side effects of a flush.

Thread 1: Make a flushless context A
Thread 1: Make another flushless context, shared
Thread 2: Make a context, shared (don't care about flushless)
Thread 1: Make a renderbuffer.
Thread 1: glClear() it to green.
Thread 1: glFinish()
Thread 1: glClear() it to red.
Thread 2: Do a glReadPixels() and make sure it's green, otherwise skip
Thread 1: Switch to the other flushless context.
Thread 2: sleep(.5)
Thread 2: Make sure glReadPixels() is still green, otherwise fail.

On multi-queue hardware, the implementation could maybe have flushed but
not completed by .5s, but I think this will be way more reliable than
comparing FPS.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/piglit/attachments/20141012/2ec9ab4a/attachment.sig>