https://bugs.documentfoundation.org/show_bug.cgi?id=144598
--- Comment #16 from Luboš Luňák <[email protected]> --- (In reply to Julien Nabet from comment #5) > should we use "Disable hardware acceleration" to set "UseSkia" at False ? No. Skia/Raster has nothing to do with hardware acceleration. > Heiko/Xisco/Luboš: Should we create a new entry like 'Disable Skia' or I don't quite see the point, we generally do not provide options to disable things that are expected to work. It makes sense to disable HW acceleration, because that depends on HW drivers, which are outside of our control. But 'Disable Skia' would be like 'Disable Gtk' or 'Disable Fontconfig', we don't provide options to disable fontconfig just because there may be a bug in it. > Luboš: I read in vcl/skia/README that "raster" was "CPU-based drawing (here > primarily used for debugging)" That's misleading, Raster is also used as a fallback when Vulkan is not available. (In reply to Mike Kaganski from comment #15) > Note e.g. vroundps, which is specific for AVX, while our baseline for > WindowsX64 is SSE2. > > This shows that my idea about the need to disable Skia/Raster is wrong, and > simply the bug around building Skia should be fixed. Or maybe it is already > fixed in a recent 7.2 version? I'm not aware of any difference here between 7.1 and 7.2, they both build SkRect using normal Clang settings which should target SSE2. But SkRect::round() is inline in the class, so it's possible there's one copy emitted built with AVX and then we get unlucky when the linker selects that one copy when merging. Since SkOpts_avx.cpp indirectly pulls in SkRect.h, that's most likely what's happening. I'll ask on the Skia mailing list. -- You are receiving this mail because: You are the assignee for the bug.
