After marking buffers as resident, Nvidia's driver seems to take a
slow path. To workaround this issue, copy to a STREAM_READ buffer and
then call GetNamedBufferSubData on it.
This is a temporary solution until we have asynchronous flushing.
Make stream buffer and cached buffers as resident and query their
address. This allows us to use GPU addresses for several proprietary
Nvidia extensions.
Instead of using as template argument a shared pointer, use the
underlying type and manage shared pointers explicitly. This can make
removing shared pointers from the cache more easy.
While we are at it, make some misc style changes and general
improvements (like insert_or_assign instead of operator[] + operator=).
Avoid copying to a staging buffer on non-granular memory addresses.
Add a callable argument to StreamBufferUpload to be able to copy to the
staging buffer directly from ReadBlockUnsafe.
Drop the std::list hack to allocate memory indefinitely.
Instead use a custom allocator that keeps references valid until
destruction. This allocates fixed chunks of memory and puts pointers in
a free list. When an allocation is no longer used put it back to the
free list, this doesn't heap allocate because std::vector doesn't change
the capacity. If the free list is empty, allocate a new chunk.
Most overlaps in the buffer cache only contain one mapped address.
We can avoid close to all heap allocations once the buffer cache is
warmed up by using a small_vector with a stack size of one.
Instead of using boost::icl::interval_map for caching, use
boost::intrusive::set. interval_map is intended as a container where the
keys can overlap with one another; we don't need this for caching
buffers and a std::set-like data structure that allows us to search with
lower_bound is enough.
Reduces some header churn and reduces rebuilds when some header
internals change.
While we're at it we can also resolve a missing include in buffer_cache.
Drop MemoryBarrier from the buffer cache and use Maxwell3D's register
WaitForIdle.
To implement this on OpenGL we just call glMemoryBarrier with the
necessary bits.
Vulkan lacks this synchronization primitive, so we set an event and
immediately wait for it. This is not a pretty solution, but it's what
Vulkan can do without submitting the current command buffer to the queue
(which ends up being more expensive on the CPU).
The original idea of returning pointers is that handles can be moved.
The problem is that the implementation didn't take that in mind and made
everything harder to work with. This commit drops pointer to handles and
returns the handles themselves. While it is still true that handles can
be invalidated, this way we get an old handle instead of a dangling
pointer.
This problem can be solved in the future with sparse buffers.
Delay buffer destruction some extra frames to avoid destroying buffers
that are still being used from older frames. This happens on Nvidia's
driver with mailbox.
`boost::make_iterator_range` is available when `boost/range/iterator_range.hpp` is included.
Also include `boost/icl/interval_map.hpp` and `boost/icl/interval_set.hpp`.
Nvidia's OpenGL driver maps gl(Named)BufferSubData with some requirements
to a fast. This path has an extra memcpy but updates the buffer without
orphaning or waiting for previous calls. It can be seen as a better
model for "push constants" that can upload a whole UBO instead of 256
bytes.
This path has some requirements established here:
http://on-demand.gputechconf.com/gtc/2014/presentations/S4379-opengl-44-scene-rendering-techniques.pdf#page=24
Instead of using the stream buffer, this commits moves constant buffers
uploads to calls of glNamedBufferSubData and from my testing it brings a
performance improvement. This is disabled when the vendor is not Nvidia
since it brings performance regressions.
