Bug 100622

Created attachment 130759 [details]
Test case which clears a depth-stencil attachment

Attached: test application to reproduce bug (see https://github.com/rikusalminen/querystall for build scripts, etc. needs c99 compiler and CFLAGS).

Clearing a depth-stencil buffer when occlusion query (GL_SAMPLES_PASSED) is enabled causes the following GL debug message (via glDebugMessageCallback):
GL_DEBUG_SOURCE_API GL_DEBUG_TYPE_PERFORMANCE GL_DEBUG_SEVERITY_MEDIUM  id: 0xb
CPU mapping a busy query object BO stalled and took 2.003 ms.

Clearing GL_DEPTH_COMPONENT24 target (instead of GL_DEPTH24_STENCIL8) does not cause this behavior.

GL_TIME_ELAPSED or GL_PRIMITIVES_GENERATED queries do not give this error.

Here are some helpful notes from #intel-gfx IRC log:

21:12 < Kayden> ugh.  the query code is fine.
21:13 < Kayden> you're hitting lunacy in meta
21:13 < Kayden> https://cgit.freedesktop.org/mesa/mesa/commit/?h=bb77745681e179b53ea64da7bc1987b57643b7b0
21:14 < Kayden> the clear isn't hitting our best-case scenario, and is falling back to a meta use-GL-to-implement-GL path, that wants to disable occlusion queries so our internal usage doesn't contribute fragments
21:14 < Kayden> and the way it does that stalls
21:14 < Kayden> we've been killing off this awful code for a while now, but some bits are still left
21:21 < Kayden> wow, that's in a sorrier state than I thought :/
21:21 < Kayden> jekstrand: it looks like we always fall through to meta for stencil clears
21:21 < Kayden> blorp only does color...
21:21 < Kayden> fast depth only does depth...
21:22 < Kayden> I don't know why we wouldn't be using blorp for stencil
21:23 < Kayden> erm
21:23 < Kayden> jekstrand: it looks like the GL driver is never using  
                blorp_clear_depth_stencil