The "VK" prefix predates the "Vulkan" namespace. It was carried around
the codebase for consistency. "VKDevice" currently is a bad alias with
"VkDevice" (only an upcase character of difference) that can cause
confusion. Rename all instances of it.
The current texture cache has several points that hurt maintainability
and performance. It's easy to break unrelated parts of the cache
when doing minor changes. The cache can easily forget valuable
information about the cached textures by CPU writes or simply by its
normal usage.The current texture cache has several points that hurt
maintainability and performance. It's easy to break unrelated parts
of the cache when doing minor changes. The cache can easily forget
valuable information about the cached textures by CPU writes or simply
by its normal usage.
This commit aims to address those issues.
This reworks how host<->device synchronization works on the Vulkan
backend. Instead of "protecting" resources with a fence and signalling
these as free when the fence is known to be signalled by the host GPU,
use timeline semaphores.
Vulkan timeline semaphores allow use to work on a subset of D3D12
fences. As far as we are concerned, timeline semaphores are a value set
by the host or the device that can be waited by either of them.
Taking advantange of this, we can have a monolithically increasing
atomic value for each submission to the graphics queue. Instead of
protecting resources with a fence, we simply store the current logical
tick (the atomic value stored in CPU memory). When we want to know if a
resource is free, it can be compared to the current GPU tick.
This greatly simplifies resource management code and the free status of
resources should have less false negatives.
To workaround bugs in validation layers, when these are attached there's
a thread waiting for timeline semaphores.
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=).
Nvidia recently introduced a new memory type for data streaming
(awesome!), but yuzu was assuming that all heaps had enough memory
for the assumed stream buffer size (256 MiB).
This worked fine on AMD but Nvidia's new memory heap was smaller than
256 MiB. This commit changes this assumption and allocates a bit less
than the size of the preferred heap, with a maximum of 256 MiB (to avoid
allocating all system memory on integrated devices).
- Fixes a crash on NVIDIA 450.82.0.0
The stream buffer before this commit once it was full (no more bytes to
write before looping) waiting for all previous operations to finish.
This was a temporary solution and had a noticeable performance penalty
in performance (from what a profiler showed).
To avoid this mark with fences usages of the stream buffer and once it
loops wait for them to be signaled. On average this will never wait.
Each fence knows where its usage finishes, resulting in a non-paged
stream buffer.
On the other side, the buffer cache is reimplemented using the generic
buffer cache. It makes use of the staging buffer pool and the new
stream buffer.
Instead of passing by copy an execution context through out the whole
Vulkan call hierarchy, use a command buffer view and fence view
approach.
This internally dereferences the command buffer or fence forcing the
user to be unable to use an outdated version of it on normal usage.
It is still possible to keep store an outdated if it is casted to
VKFence& or vk::CommandBuffer.
While changing this file, add an extra parameter for Flush and Finish to
allow releasing the fence from this calls.
This manages two kinds of streaming buffers: one for unified memory
models and one for dedicated GPUs. The first one skips the copy from the
staging buffer to the real buffer, since it creates an unified buffer.
This implementation waits for all fences to finish their operation
before "invalidating". This is suboptimal since it should allocate
another buffer or start searching from the beginning. There is room for
improvement here.
This could also handle AMD's "pinned" memory (a heap with 256 MiB) that
seems to be designed for buffer streaming.
