Full pipeline barriers between every RP can be extremely expensive on HW, by analysing the inputs and outputs of a draw it's possible to construct a much more optimal barrier that only syncs what is neccessary.
By importing guest memory directly onto the host GPU we can avoid many of the complexities that occur with memory tracking as well as the heavy performance overhead in some situations. Since it's still desired to support the traditional buffer method, as it's faster in some cases and more widely supported, most of the exposed buffer methods have been split into two variants with just a small amount of shared code. While in most cases the code is simpler, one area with more complexity is handling CPU accesses that need to be sequenced, since we don't have any place we can easily apply writes to on the GPFIFO thread that wont also impact the buffer on the GPU, to solve this, when the GPU is actively using a buffer's contents, an interval list is used to keep track of any GPFIO-written regions on the CPU and any CPU reads to them will instead be directed to a shadow of the buffer with just those writes applied. Once the GPU has finished using buffer contents the shadow can then be removed as all writes will have been done by the GPU.
The main caveat of this is that it requires tying host sync to guest sync, this can reduce performance in games which double buffer command buffers as it prevents us from fully saturating the CPU with the GPFIFO thread.
Due to the frequency at which is is called megabuffering performance is critical to the performance of the entire emulator, especially in high-drawcall-count scenarios. After the view redesign, megabuffering on a per-view level was no longer possible nor desirable, and thus megabuffering was modified to just copy for every usage of a view. This worked great at the time since there were other bottlenecks, however gpu-new has since removed almost all of them and megabuffering is now a major sore point. Fix this by megabuffering small chunks and storing them in a page-table like structure within the buffer, these chunks can be referenced by multiple views and will be smartly invalidated whenever the sequence number or execution number changes to avoid any sequencing issues. In addition to this, to help the case where almost the whole buffer is read every single frame across a set of multiple views, an optimisation to skip the chunked tracking and use one large single megabuffer allocation and one single memcpy has been introduced. This reduces the overall amount of time spent in memcpy since large memcpys are quicker.
When profiling SMO, it became obvious that the constant locking of textures and buffers in SyncDescriptors took up a large amount of CPU time (3-5%), a precious resource in intensive areas like Metro. This commit implements somewhat of a workaround to avoid constant relocking, if a buffer is frequently attached on the GPU and almost never used on the CPU we can keep the lock held between executions. Of course it's not that simple though, if the guest tries to lock a texture for the first time which has already been locked as preserve on the GPFIFO we need to avoid a deadlock. This is acheived through a combination of two things: first we periodically clear the locked attachments every 2*SlotCount submissions, preventing a complete deadlock on the CPU (just a long wait instead) and meaning that the next time the resource is attached on the GPU it will not be marked for preservation due to having been locked on the guest before; second, we always need to unlock everything when the GPU thread runs out of work, as the perioding clearing will not execute in this case which would otherwise leave the textures locked on the GPFIFO thread forever (if guest was waiting on a lock to submit work). It should be noted that we don't clear preserve attached resources in the latter scenario, only unlock them and then relock when more work is available.
For the upcoming preserve attachment optimisation, which will keep buffers/textures locked on the GPU between executions, we don't want to preserve any which are frequently locked on the CPU as that would result in lots of needless waiting for a resource to be unlocked by the GPU when it occasionally frees all preserve attachments when it could have been done much sooner. By checking if a resource has ever been locked on the CPU and using that to choose whether we preserve it we can avoid such waiting.
When a buffer is trapped nearly every frame, the cost of trapping and synchronising its contents starts to quickly add up. By always using the megabuffer when this is the case, since megabuffer copies are done directly from the guest, we skip the need to synchronise/trap the backing.
The original intention was to cache on the user side, but especially with shader constant buffers that's difficult and costly. Instead we can cache on the buffer side, with a page-table like structure to hold variable sized allocations indexed by the aligned view base address. This avoids most redundant copies from repeated use of the same buffer without updates inbetween.
This isn't a guarantee provided by actual HW so we don't need to provide it either, the sync can be skipped once the buffer already been synced at least once within the execution.
Constructing the GPU copy callback in `ConstantBuffers::Load()` ended up taking a fair amount of time despite it almost never being used in practice. By making it optional it can be skipped most of the time and only done when it's actually neccessary by calling `Write()` again if the initial call returned true.
Buffer views creation was a significant pain point, requiring several layers of caching to reduce the number of creations that introduced a lot of complexity. By reworking delegates to be per-buffer rather than per-view and then linearly allocating delegates (without ever freeing) views can be reduced to just {delegatePtr, offset, size}, avoiding the need for any allocations or set operations in GetView. The one difficulty with this is the need to support buffer recreation, which is achived by allowing delegates to be chained - during recreation all source buffers have their delegates modified to point to the newly created buffer's delegate. Upon accessing a view with such a chained delegate the view will be modified to point directly to the end delegate with offset being updated accordingly, skipping the need to traverse the chain for future accesses.
Currently we heavily thrash the heap each draw, with malloc/free taking up about 10% of GPFIFOs execution time. Using a linear allocator for the main offenders of buffer usage callbacks and index/vertex state helps to reduce this to about 4%
After the introduction of workahead a system to hold a single large megabuffer per submission was implemented, this worked fine for most cases however when many submissions were flight at the same time memory usage would increase dramatically due to the amount of megabuffers needed. Since only one megabuffer was allowed per execution, it forced the buffer to be fairly large in order to accomodate the upper-bound, even further increasing memory usage.
This commit implements a system to fix the memory usage issue described above by allowing multiple megabuffers to be allocated per execution, as well as reuse across executions. Allocations now go through a global allocator object which chooses which chunk to allocate into on a per-allocation scale, if all are in use by the GPU another chunk will be allocated, that can then be reused for future allocations too. This reduces Hollow Knight megabuffer memory usage by a factor 4 and SMO by even more.
The code is much simpler to reason about when reading the code as it doesn't require evaluating all the potential edge cases of trap handlers in different states. It should be noted that this should not change behavior in any meaningful way, at most it can prevent a minor race where the protection could be upgraded after being downgraded by the signal handler leading to a redundant trap.
The `TrapRegions` function performed a page-out on any regions that were trapped as read-only, this wasn't optimal as it would tie them both into the same operation while Buffers/Textures require to protect then synchronize and page-out. The trap was being moved to after the synchronize to get around this limitation but that can cause a potential race due to certain writes being done after the synchronization but prior to the trap which would be lost. This commit fixes these issues by splitting paging out into `PageOutRegions` which can be called after `TrapRegions` by any API users.
Co-authored-by: Billy Laws <blaws05@gmail.com>
`NCE::TrapRegions` was a bit too overloaded as a method as it implicitly trapped which was unnecessary in all current usage cases, this has now been made more explicit by consolidating the functionality into `NCE::CreateTrap` which handles just creation of the trap and nothing past that, `RetrapRegions` has been renamed to `TrapRegions` and handles all trapping now.
Co-authored-by: Billy Laws <blaws05@gmail.com>
Having a single variable denoting the exact state of a buffer and the operations that could be performed on it was found to be too restrictive, it's now been expanded into an additional `BackingImmutability` variable but due to these two. We can no longer use atomics without significant additional complexity so all accesses to the state are now mediated through `stateMutex`, a mutex specifically designed for tracking the state.
While designing the system around `stateMutex` it was determined to be more efficient than atomics as it would enforce blocking far less than it would generally have been compared to if the regular atomic fallback of locking the main resource lock which is locked for significantly longer generally.
Co-authored-by: PixelyIon <pixelyion@protonmail.com>
As a performance sensitive part of code, the NCE Trapping API benefits from having tracing and it helps us better determine where guest code is spending its time for more targeted optimizations.
The lifetime of the `this` pointer in the trap callbacks could be invalid as the lifetime of the underlying `Buffer`/`Texture` object wasn't guaranteed, this commit fixes that by passing a `weak_ptr` of the objects into the callbacks which is locked during the callbacks and ensures that a destroyed object isn't accessed.
Co-authored-by: Billy Laws <blaws05@gmail.com>
It was determined that `FindOrCreate` has several issues which this commit fixes:
* It wouldn't correctly handle locking of the newly created `Buffer` as the constructor would setup traps prior to being able to lock it which could lead to UB
* It wouldn't propagate the `usedByContext`/`everHadInlineUpdate` flags correctly
* It wouldn't correctly set the `dirtyState` of the buffer according to that of its source buffers
The condition for `setDirty` in the dirty state CAS was inverted from what it should've been resulting in synchronizing incorrectly, this commit fixes the condition to correct synchronization.
`ContextLock` had unoptimal semantics in the form of direct access to the `isFirst` member which wasn't clearly defined, it's now been broken up into function calls `IsFirstUsage` and `OwnsLock` with explicit move semantics and a function for releasing the lock.
Co-authored-by: PixelyIon <pixelyion@protonmail.com>
The buffer's non-blocking behavior could lead to an invalid state where the dirty state doesn't adequately represent the buffer's true state, the check has now been moved inside the CAS loop as its behavior changes depending on the dirty state. In addition, `SynchronizeGuest` returns a boolean denoting if the synchronization was successful now to make code flows depending on non-blocking synchronization cleaner.
`SynchronizeGuest` could only set the dirty state to `Clean` which was redundant since calls to it from inside the write trap handler would set it to `CpuDirty` directly after, this fixes that by doing it inside the function when necessary.
If a `FenceCycle` isn't attached then `PollFence` returned `false` while it should return if the buffer has any concurrent GPU usages in flight, this has now been fixed by returning `true` in those cases.
The GPU inline copy callback was broken for `Buffer::Write` as it wasn't always called when it needed to be and didn't handle attaching of the buffer to the executor which would cause it to be unlocked. This commit addresses both of these issues, it introduces a `AttachLockedBuffer` method to attach an already locked buffer to the executor.
A deadlock was caused by holding `trapMutex` while waiting on the lock of a resource inside a callback while another thread holding the resource's mutex waits on `trapMutex`. This has been fixed by no longer allowing blocking locks inside the callbacks and introducing a separate callback for locking the resource which is done after unlocking the `trapMutex` which can then be locked by any contending threads.
We generally don't need to lock the `Texture`/`Buffer` in the trap handler, this is particularly problematic now as we hold the lock for the duration of a submission of any workloads. This leads to a large amount of contention for the lock and stalling in the signal handler when the resource may be `Clean` and can simply be switched over to `CpuDirty` without locking and utilizing atomics which is what this commit addresses.
We utilized a `FenceCycle` to keep track of if the buffer was mutable or not and introduced another cycle to track GPU-side requirements only on fulfillment of which could the buffer be utilized on the host but due to the recent change in the behavior this system ended up being unoptimal.
This commit replaces the cycle with a boolean tracking if there are any usages of the resource on the GPU within the current context that may prevent it from being mutated on the CPU. The fence of the context is simply attached to the buffer based off this which was allowed as the new behavior of buffer fences matches all the requirements for this.
An atomic transactional loop was performed on the backing `std::shared_ptr` inside `BufferView`/`TextureView`'s `lock`/`LockWithTag`/`try_lock` functions, these locks utilized `std::atomic_load` for atomically loading the value from the `shared_ptr` recursively till it was the same value pre/post-locking.
This commit abstracts the locking functionality of `TextureView`/`BufferDelegate` into `LockableSharedPtr` to avoid code duplication and removes the usage of `std::atomic_load` in either case as it is not necessary due to the implicit memory barrier provided by locking a mutex.
GPU resources have been designed with locking by fences in mind, fences were treated as implicit locks on a GPU, design paradigms such as `GraphicsContext` simply unlocking the texture mutex after attaching it which would set the fence cycle were considered fine prior but are unoptimal as it enforces that a `FenceCycle` effectively ensures exclusivity. This conflates the function of a mutex which is mutual exclusion and that of the fence which is to track GPU-side completion and led to tying if it was acceptable to use a GPU resource to GPU completion rather than simply if it was not currently being used by the CPU which is the function of the mutex.
This rework fixes this with the groundwork that has been laid with previous commits, as `Context` semantics are utilized to move back to using mutexes for locking of resources and tracking the usage on the GPU in a cleaner way rather than arbitrary fence comparisons. This also leads to cleaning up a lot of methods that involved usage of fences that no longer require it and therefore can be entirely removed, further cleaning up the codebase. It also opens the door for future improvements such as the removal of `hostImmutableCycle` and replacing them with better solutions, the implementation of which is broken at the moment regardless.
While moving to `Context`-based locking the question of multiple GPU workloads being in-flight while using overlapping resources came up which brought a fundamental limitation of `FenceCycle` to light which was that only one resource could be concurrently attached to a cycle and it could not adequately represent multi-cycle dependencies. `FenceCycle` chaining was designed to fix this inadequacy and allows for several different GPU workloads to be in-flight concurrently while utilizing the same resources as long as they can ensure GPU-GPU synchronization.
If we want to allow submitting multiple pieces of work to the GPU at once while still requiring CPU synchronization, we'll need to track all past fence cycles associated with a resource alongside the current one. To solve this the concept of chaining fences has been introduced, fences from past usages can be chained to the latest fence which'll then recursively forward operations to chained fences.
This change also ends up mandating a move away from `FenceCycleDependency` as it would prevent fences from concurrently locking the same resources which is required for chaining to work as two fences being chained fundamentally means they're locking the same resources. The `AtomicForwardList` is therefore used as the new container.
