The constant destruction and creation of `BufferView`s in cbuf-heavy games showed up as a large chunk of the profiler. Fix this by taking advantage of the fact that constant buffer `BufferView`s are never deleted and always kept around in the cache to just return a pointer to them in the cache.
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%
Certain titles can have a display frames out of order due to not waiting on the copy from the final RT to the swapchain image to occur. Although `PresentFrame` does wait on the syncpoint, that isn't enough to ensure the source texture is up-to-date due to us signalling syncpoints early.
By waiting on the swapchain texture after the copy is submitted, we now implicitly wait on the source texture's cycle to be signalled thus waiting on the frame to be done which fixes the issue.
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.
Accesses to unset entries is now clearly defined as returning a 0'd out value, the prior behavior would be to optimize sets for border segments to use L2 atomicity when the specific segment had no L1 entries set. This would lead to any future lookups of offsets within the same L2 segment but a different L1 entry to incorrectly return an inaccurate value as the only prior guarantee was that lookups after setting a segment would return the same value as was set but lacked the guarantee for unset segments to also consistently return unset values.
This could lead to issues in practical usages such as the `BufferManager` lookups returning the existence of a `Buffer` at a location falsely even though the segment was never set to the value, this was problematic as raw pointers were utilized and bound checks would lead to a segmentation fault.
This commit fixes this issue by introducing this guarantee and refactoring the class accordingly, it also deletes the `Set` method for setting a single entry as the meaning is ambiguous and it's functionality was more akin to the past guarantee and no longer makes sense.
Co-authored-by: PixelyIon <pixelyion@protonmail.com>
We would always write all L1 entries that correspond to an L2 entry, even if setting an input range ended before that. This would effectively reduce the atomicity of the segment table to that of the L2 range and lead to breaking API guarantees by returning entirely wrong segment values for a lookup covering a region that was overwritten.
It was determined that `RangeTable` was too ambiguous of a name as it could be interpreted to be holding ranges rather than looking them up, to avoid confusion the terminology has been changed to `range` to `segment`. As "segment table" is more clear in describing that it is a table comprised of descriptors regarding segments and it avoids any overlaps with terminology concerning "pages" which would be overly specific for this data structure or the ambiguous "ranges".
The PI CAS in `MutexUnlock` ends up loading `basePriority` rather than `priority` which could lead to an infinite CAS loop when `basePriority` doesn't equal to `priority` and the `highestPriorityThread`'s priority is lower than `basePriority`.
It was determined that `Texture::SynchronizeGuest`'s `TextureBufferCopy` had races that were exposed by the introduction of the cycle waiter thread, the synchronization did not take place under a locked context so the texture could be mutated at any point in addition to the destructor not being run during `FenceCycle::Wait` due to `shouldDestroy` being `false`.
This commit fixes the issue by making `SynchronizeGuest` entirely blocking as all usages of the function required blocking semantics regardless so it would be pointless to retain its async nature while solving any races that may arise from it being async.
Co-authored-by: Billy Laws <blaws05@gmail.com>
Since we don't call `SynchronizeHost` on source buffers which are GPU dirty, their mirrors will be out of date. The backing contents of this source buffer's region in the new buffer will be incorrect. By copying from the backing directly, we can ensure that no writes are lost and that if the newly created buffer needs to turn GPU dirty during recreation no copies need to be done since the backing is as up to date as the mirror at a minimum.
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.
Two issues exist with locking of `KThread::waiterMutex`:
* It was not always locked when accessing waiter members such as `waitThread`, `waitKey` and `waitTag` which would lead to a race that could end up in a deadlock or most notably a segfault inside `UpdatePriorityInheritance`
* There could be a deadlock from `UpdatePriorityInheritance` locking `waiterMutex` of a thread and waiting to get the owner's `waiterMutex` while on another thread `MutexUnlock` holds the owner's `waiterMutex` and waits on locking the `waiterMutex` held by `UpdatePriorityInheritance`
This commit fixes both issues by adding appropriate locking to all locations where waiter members are accessed in addition to adding a fallback mechanism inside `UpdatePriorityInheritance` that unlocks `waiterMutex` on contention to avoid a deadlock.
The condition for exiting the CAS loops is incorrect in several places which leads to additional loops, while this doesn't make the behavior incorrect it does lead to redundant iterations.
Co-authored-by: Billy Laws <blaws05@gmail.com>
A substantial amount of time would be spent on creation/destruction of `VkDescriptorSet` which scales on titles doing a substantial amount of draws with bindings, this leads to poor performance on those titles as the frametime is dragged down by performing these tasks while they repeatedly create descriptor sets of the same layouts.
This commit fixes it by pooling descriptor sets per-layout in a dynamically resizable pool and keeping them around rather than destroying them after usage which leads to the vast majority of cases not requiring a new descriptor set to even be created. It leads to significantly improved performance where it would otherwise be spent on redundant destruction/recreation or push descriptor updates which took a substantial amount of time themselves.
Additionally, the `BaseDescriptorSizes` were not kept up to date with all of the descriptor types, it led to no crashes on Adreno/Mali as they were purely used for size calculations on either driver but has been corrected to avoid any future issues.
A substantial amount of time is spent destroying dependencies for any threads waiting or polling `FenceCycle`s, this is not optimal as it blocks them from moving onto other tasks while destruction is a fundamentally async task and can be delayed.
This commit solves this by introducing a thread that is dedicated to waiting on every `FenceCycle` then signalling and destroying all dependencies which entirely fixes the issue of destruction blocking on more important threads.
Buffer lookups are a fairly expensive operation that we currently spend `O(log n)` on the simplest and most frequent case of which is a direct match, this is a very frequent operation where that may be insufficient. This commit optimizes that case to `O(1)` by utilizing a `RangeTable` at the cost of slightly higher insertion/deletion costs for setting ranges of values but these are minimal in frequency compared to lookups.
A data structure that can represent the same value for a range of addresses (pages) is required for fast lookup in certain cases. This commit implements a near optimal data structure for mass insertion and O(1) lookup of range-based data, this is achieved using the host MMU and implementing multiple levels of atomicity for the ranges.
It should be noted that the table is limited to two levels but can be extended to a variable amount of ranges in the future, it was determined that additional levels of ranges can be beneficial for performance depending on the specific use-case.
Adreno drivers have certain errata which leads to Vulkan Push Descriptors to be broken on them in certain cases which leads to a descriptor set update being swallowed. This has been worked around by disabling push descriptors on Adreno drivers, this may lead to reduced performance on certain titles which frequently bind new descriptors.
Any semaphore releases are implicit synchronization events that can be utilized by the guest to pick up that the GPU has executed till a certain point and therefore we must submit all prior work accordingly.
DMA copies utilized `SubmitWithFlush` instead of `Submit`, this is not required and incurs significant additional synchronization penalties which will no longer be required.
We want to avoid blocking on surface creation unless necessary, this commit doesn't wait on the creation of the surface as it default initializes the value which'll generally be `Identity` or the transformation of the previous surface if it was lost.
Co-authored-by: Billy Laws <blaws05@gmail.com>
The V-Sync `KEvent` would be used by the presentation thread prior to construction leading to dereferencing an invalid value, this has been fixed by changing the order of construction to move the construction of the presentation thread after the V-Sync event.
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>
Similar to `Buffer`s, `Texture`s suffered from unoptimal behavior due to using atomics for `DirtyState` and had certain bugs with replacement of the variable at times where it shouldn't be possible which have now been fixed by moving to using a mutex instead of atomics. This commit also updates the API to more closely match what is expected of it now and removes any functions that weren't utilized such as `SynchronizeGuestWithBuffer`.
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>
The `CommandExecutor`'s `MegaBuffer` was not being updated with the latest `FenceCycle` on being flushed in `SubmitWIthFlush`, this led to the megabuffer being overwritten prior to its GPU-side usage being complete. This commit fixes that by replacing the cycle to the latest cycle and prevents any races that occurred prior.
`FindOrCreate` ended up being monolithic function with poor readability, this commit addresses those concerns by refactoring the function to split it up into multiple member functions of `BufferManager`, while some of these member functions may only have a single call-site they are important to logically categorize tasks into individual functions. The end result is far neater logic which is far more readable and slightly better optimized by virtue of being abstracted better.
In certain cases the move constructor may not suffice and the move assignment operator is required, this commit implements that and moves to using a pointer for storing the `resource` member rather than a reference as its semantics matched what we desired more and allowed for assignment of the `resource`.
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.
The formats of the textures involved in a texture were checked for equality, this broke certain copies as the presentation engine would invoke copies between textures of different yet compatible formats.
Co-authored-by: PixelyIon <pixelyion@protonmail.com>
`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 position at which we call submit is a significant factor in performance and we did so at the end of PBs (PushBuffers), this isn't optimal as there could be multiple PBs queued up that would benefit from being in the same submission. We now delay the submission of the workload till we run out of PBs.
A buffer that's attached to a context could be coalesced into a larger buffer which isn't attached, this would break as it wouldn't keep the buffer alive till the end of the associated context. To fix this if any source buffers are attached then the resulting coalesced buffer is also attached now.
The CAS condition for KThread PI was inverted which lead to entirely incorrect behavior for CAS conditions which while it might work in the vast majority of cases would lead to significantly inaccurate behavior.
The lock callback would `continue` which would end up skipping over the current item as it applied to the inner loop rather than the outer loop as intended. This has now been fixed by using `break` and a check instead.
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.
The trap callbacks did not wait on the `Texture` to complete synchronization to the guest, this resulted in races where the contents written to the texture would be overwritten by the synced content. This commit fixes that by waiting on the fences at the end of the trap callback.
