This means SwiftShader is only used for Mac ES3 build and Android.
Also remove the parallelism and --output-on-failure, it's unhelpful as
it only leads to less information being provided if the CI fails.
This apparently works well for quite some time (this is Ubuntu 18.04,
after all!), I just didn't know because the build defaults to GLX which
doesn't work headless. So, yeah, code coverage, here I come!
This I like, a notification sufficiently in advance, that a certain
version of an image is deprecated. Not the whole OS version altogether,
not the platform as a whole.
It fails on a linker error otherwise. There's a bit of annoying logic
needed for pre-GLVND systems as there it's not possible to link to GLX
without dragging the whole libGL in as well, causing bad conflicts with
libGLES. Which means I can't test this on the CI yet as there CMake is
forced to version 3.5 and finding GLVND is only in 3.10.
Interesting that neither GlxApplication nor XEglApplication was actually
built on the CI. Also test all EGL applications and contexts on GLX
builds as well, as those should work there too. For GLX on EGL builds
it's a different story.
It compiles on GLES2 as well, but there it hits the massive PITA of
being unable to render to LUMINANXCE formats and GL_RED formats not
really being available everywhere.
I don't have the patience to fix that, and almost nobody needs to use
ES2 platforms nowadays, so this isn't really a priority.
Except the base GL and Vulkan one. Because usually the nondeprecated
builds needs the most iterations and any builds that run after it failed
are just wasting credits.
Take the linux-gl, linux-vulkan and linux-nondeprecated builds as the
main sanity checks. If they pass, then the rest gets scheduled as well,
with further dependencies between them (such as WebGL2 tests being run
only if GLES3 passes).
Otherwise it often happens that until I'm able to push a fix for GCC
4.8 or an ES2-specific build failure, it already consumes 10 minutes on
15 jobs, which is quite a lot, especially for the macOS jobs.
FUCKING stupid defaults. CMake's default no less stupid. I'm MAD, why
would anybody even want to have the build run sequentially on todays
machines?! If your build is crap and can't run in parallel, FIX YOUR
CODE, but don't make the other 99% users suffer!!
It's just not worth the pain. I have to reduce the parallelism from 24
to 6, and yet it still sometimes gets stuck. And at that point it's
about twice as slow, so there's not really any gain from this costing
half the credits.
This reverts commit 9d61a63553.
This reverts commit 9c4f2ceea2.
This reverts commit 80b7694468.
I'm running out of the 400k free monthly credits and since macOS takes
50 credits per minute vs 5 for the Linux Docker, it's the obvious
candidate. ARM and Android take 10 instead of 5 and since we don't have
any ARM-specific code and the Android ES3 build doesn't run GL tests,
it's also non-essential.
Will revert this at the start of Junly.
The main part of the build time is fetching packages (where I doubt
multithreading can help anything) and running tests, which is serial to
have the output reasonably ordered. The actual build time takes less
than a half of the total time so going with the smallest resource class
instead of the medium shouldn't make that much of a difference. But it
uses half the credits, and that's what matters.
Vladimír Vondruš