Author :
Charlie Lao
Date :
2023-12-15 10:17:32
Hash :40f4de8f Message :Vulkan: Ensure we use cached memory for readPixels stagingBuffer
Previous CL crrev.com/c/5112759 does not solve the performance issue for
ChromeOS. The reason is that on more recent intel GPU, there is no
hostVisibleCachedCoherent heap. When we allocate staging buffer, we
specify CachedCoherent as the preferredFlags instead of requiredFlags.
This means we still end up getting UncachedCoherent since VMA tries to
respect coherent bits as first priority. This CL Changes CachedCoherent
to CachedPreferCoherent, and made Cached as required bit, thus ensures
the memory allocated is cached. Since coherent bit may not be honored,
thus we have to call invalidate/flush (which underline implementation
will check the bit and early out if no need).
Somehow on ARM GPU using cachedNonCoherent staging buffer causing many
test failures, even though we do call invalidate() after allocation, and
tests pass on all other GPUs. It almost indicates ARM driver have a bug
with invalidate() that it is not doing expected. But before I can be
sure and fixed, I added feature bit to keep ARM the old behavior, which
uses UnCached memory for readPixels which should suffer the performance
as well.
Bug: b/315836169
Bug: b/310701311
Change-Id: I1eec6105ce74275faa893b0206be8470f0cde72f
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/5122318
Commit-Queue: Charlie Lao <cclao@google.com>
Auto-Submit: Charlie Lao <cclao@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>