First, issue glMemoryBarrier() only if the read-after-write or
write-after-write occurs for a write involving incoherent memory
access. In practice this means a QRhiTexture with UsedWithLoadStore
or a QRhiBuffer used as an SSBO. Doing the barrier call for other
types of texture or buffers is pointless, and potentially
wasteful. What is more, some OpenGL implementations feature bugs that
manifest themselves when glMemoryBarrier() is called, see associated
Jira issue for details.
Second, reduce GL_ALL_BARRIER_BITS to the ones relevant for the type
of the resource, in an attempt to provide a minor optimization. Real
fine-grained resource usage tracking is not implemented for OpenGL so
this still means specifying all buffer-related or all texture-related
barriers together, but this is still better than all. As the usage of
compute is quite close to zero in Qt at the moment, this won't make
much of a difference in practice.
Fixes: QTBUG-96322
Pick-to: 6.2
Change-Id: Ibf794e2e8a4cbc3c69acaa88476e85de6646b605
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
fea1d57c2b