We have a conversion from LDG on the compute shader to a special constant buffer binding that's used to exceed hardware limits on compute, but it was only running if the byte offset could be identified. The fallback that checks all of the bindings at runtime only checks the storage buffers.
This PR adds checking ube ranges to the LoadGlobal fallback. This extends the changes in #4011 to only check ube entries which are accessed by the shader.
Fixes particles affected by the wind in The Legend of Zelda: Breath of the Wild. May fix other weird issues with compute shaders in some games.
Try a bunch of games and drivers to make sure they don't blow up loading constants willynilly from searchable buffers.
* Initial implementation of metal surface across UIs
* Fix SDL2 on windows
* Update Ryujinx/Ryujinx.csproj
Co-authored-by: Mary-nyan <thog@protonmail.com>
* Address Feedback
Co-authored-by: Mary-nyan <thog@protonmail.com>
* GPU: Use lazy checks for specialization state
This PR adds a new class, the SpecializationStateUpdater, that allows elements of specialization state to be updated individually, and signal the state is checked when it changes between draws, instead of building and checking it on every draw. This also avoids building spec state when
Most state updates have been moved behind the shader state update, so that their specialization state updates make it in before shaders are fetched.
Downside: Fields in GpuChannelGraphicsState are no longer readonly. To counteract copies that might be caused this I pass it as `ref` when possible, though maybe `in` would be better? Not really sure about the quirks of `in` and the difference probably won't show on a benchmark.
The result is around 2 extra FPS on SMO in the usual spot. Not much right now, but it will remove costs when we're doing more expensive specialization checks, such as fragment output type specialization for macos. It may also help more on other games with more draws.
* Address Feedback
* Oops
* GPU: Swap bindings array instead of copying
Reduces work on UpdateShaderState. Now the cost is a few reference moves for arrays, rather than copying data.
Downside: bindings arrays are no longer readonly.
* Micro optimisation
* Add missing docs
* Address Feedback
* Track buffer migrations and flush source on incomplete copy
Makes sure that the modified range list is always from the latest iteration of the buffer, and flushes earlier iterations of a buffer if the data has not been migrated yet.
* Cleanup 1
* Reduce cost for redundant signal checks on Vulkan
* Only inherit the range list if there are pending ranges.
* Fix OpenGL
* Address Feedback
* Whoops
Polygon topology wasn't really supported and would only work on OpenGL on drivers that haven't removed it. As an alternative, this PR makes all cases of polygon topology use triangle fan. The topology type and transform feedback type have not been changed, as I don't think geo shader/tfb should be used with polygons.
The OpenGL spec states:
Only convex polygons are guaranteed to be drawn correctly by the GL.
For convex polygons, triangle fan is equivalent to polygon. I imagine this is probably how it works on device, as this get-out-of-jail-free card is too enticing to pass up.
This fixes the stat display in Pokemon S/V.
This replacement is meant to be done with the original identified byteOffset, not the one assigned later on by the below conditionals (that already has the constant offset added, for instance).
This fixes videos being pixelated in Xenoblade 3, and other regressions that might have happened since #3847.
* GAL: Send all buffer assignments at once rather than individually
The `(int first, BufferRange[] ranges)` method call has very significant performance implications when the bindings are spread out, which they generally always are in Vulkan. This change makes it so that these methods are only called a maximum of one time per draw.
Significantly improves GPU thread performance in Pokemon Scarlet/Violet.
* Address Feedback
Removed SetUniformBuffers(int first, ReadOnlySpan<BufferRange> buffers)
* GPU: Access non-prefetch command buffers directly
Saves allocating new arrays for them constantly - they can be quite small so it can be very wasteful. About 0.4% of GPU thread in SMO, but was a bit higher in S/V when I checked.
Assumes that non-prefetch command buffers won't be randomly clobbered before they finish executing, though that's probably a safe bet.
* Small change while I'm here
* Address feedback
I did this on ncbuffer2 when we were using it for LDN 3, but I noticed that it can apply to the current buffer manager too, and it's an easy performance win.
The only buffer access that can come from another thread is the overlap search for buffers that have been unmapped. Everything else, including modifications, come from the main GPU thread. That means we only need to lock the range list when it's being modified, as that's the only time where we'll cause a race with the unmapped handler.
This has a significant performance improvements in situations where FIFO is high, like the other two PRs. Joined together they give a nice boost (73.6 master -> 79 -> 83 fps in SMO).
A quick fix to prevent reading the wrong value of Count when reregistering ranges for a new target buffer. Buffer flushes from another thread can modify the range list when the lock isn't active, which can change the count.
This prevents some crashes in Pokemon Scarlet/Violet. It's probably likely that buffer migration during flush is causing some other issues in this game, but this at least prevents the crashing.
* Vulkan: Don't create preload buffer outside a render pass
The preload command buffer is used to avoid render pass splits and barriers when updating buffer data. However, when a render pass is not active (for example, at the start of a pass, or during compute invocations) buffer uploads can be performed at any time, so the optimization isn't as useful.
This PR makes it so that the preload command buffer is only used for buffer updates outside of a render pass. It's still used for textures as I don't want to shake things up right now regarding how the preload buffer is obtained before some other changes, and texture updates are a lot rarer anyways.
Improves performance slightly in Pokemon Scarlet/Violet (43 -> 48), as it was switching to compute, writing a bunch of buffers inline, then dispatching, then flushing commands... It uses 1 command buffer instead of 2 every time it does this now. Maybe it would be nice to find a faster way to sync without creating so many command buffers in a short period of time.
* Address feedback
* Prune ForceDirty and CheckModified caches on unmap
Since we're now using this for modified checks on the HLE indirect draw method, I'm worried that leaving these to forever gather cache entries isn't the best idea for performance in the long term, and it could keep old buffer objects alive for longer than they should be.
This PR adds the ability to prune invalid entries before checking these caches, and queues it whenever gpu memory is unmapped. It also aligns modified checks to the page size, as I figured it would be possible for a huge number of overlapping over a game's runtime.
This prevents Super Mario Odyssey from having 10s of thousands of entries in the modified cache in Metro Kingdom, and them duplicating when entering and leaving a building (should be cleared, as they were unmapped).
* Address Feedback
* Allow _volatile to be set from MultiRegionHandle checks again
Tracking handles have a `_volatile` flag which indicates that the resource being tracked is modified every time it is used under a new sequence number. This is used to reduce the time spent reprotecting memory for tracking writes to commonly modified buffers, like constant buffers.
This optimisation works by detecting if a buffer is modified every time a check happens. If a buffer is checked but it is not dirty, then that data is likely not modified every sequence number, and should use memory protection for write tracking. If the opposite is the case all the time, it is faster to just assume it's dirty as we'd just be wasting time protecting the memory.
The new MultiRegionBitmap could not notify handles that they had been checked as part of the fast bitmap lookup, so bindings larger than 4096 bytes wouldn't trigger it at all. This meant that they would be subject to a ton of reprotection if they were modified often.
This does mean there are two separate sources for a _volatile set: VolatileOrDirty + _checkCount, and the bitmap check. These shouldn't interfere with each other, though.
This fixes performance regressions from #3775 in Pokemon Sword, and hopefully Yu-Gi-Oh! RUSH DUEL: Dawn of the Battle Royale. May affect other games.
* Fix stupid mistake
The type in the `texOp` in the textureSize instruction doesn't have the exact type on SPIR-V (for example, it is missing the Array flag). This PR gives it the proper type before giving it to the unscaling helper.
This fixes the ground textures being broken on Pokemon Scarlet/Violet when scaling. It wasn't finding the texture, so the descriptor index it provided was -1...
* Eliminate CB0 accesses
Still some work to do, decouple from hle?
* Forgot the important part somehow
* Fix and improve alignment test
* Address Feedback
* Remove some complexity when checking storage buffer alignment
* Update Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Due to the `using` statement being scoped to the `CreateTextureView` method, `TextureStorage` would be disposed as soon as the view was returned.
This was largely fine as the TextureStorage resources were being kept alive by the views holding their own references to them, but it also meant that dispose is only called as soon as the texture is created.
Aliased Storages are TextureStorages created with the same allocation as another TextureStorage, if they have to be aliased as another format. We keep track of a TextureStorage's `_aliasedStorages` as they are created, and dispose them when the TextureStorage is disposed...
...except it is disposed immediately, before any aliased storages are even created. The aliased storages added after this will never be disposed.
This PR attempts to fix this by disposing TextureStorage when its view count reaches 0. The other use of texture storage - the D32S8 blit - still manually disposes the storage, but regular uses created via the GAL are now disposed by the view count.
I think this makes the most sense, as otherwise in the future this behaviour might be forgotton and more things could be added to the Dispose() method that don't work due to it not actually calling at the right time.
This should improve memory leaks in Super Mario Odyssey, most noticeable when resolution scaling. The memory usage of the game is still wildly unpredictable due to how it interacts with the texture cache, but now it shouldn't get considerably longer as you play... I hope. I've seen it typically recover back to the same level occasionally, though it can spike significantly.
Please test a bunch of games on multiple GPUs to make sure this doesn't break anything.
* Fix various issues caused by #3679
- The arguments for the 0th dummy vertex buffer were incorrect - it was given an offset of 16 rather than a size of 16.
- The wrong size was used when doing `autoBuffer.Get` on a converted vertex buffer.
- The possibility of a vertex buffer being disposed and then rebound can rebindings to find a different buffer where the current range is out of bounds. Avoid binding when out of range to prevent validation errors.
- The above also affects generation of converted buffers, which was a bit more fatal. Conversion functions now attempt to bound input offset/size.
* Fix offset for converted buffer
Luigi's Mansion 3 performs a non-index quads draw with 6 vertices. It's meant to ignore the last two, but the index pattern's primitive count calculation was rounding up.
No idea why the game does this but this should fix random triangles in the map.
* GPU: Pass SpanOrArray for Texture SetData to avoid copy
Texture data is often converted before upload, meaning that an array was allocated to perform the conversion into. However, the backend SetData methods were being passed a Span of that data, and the Multithreaded layer does `ToArray()` on it so that it can be stored for later! This method can't extract the original array, so it creates a copy.
This PR changes the type passed for textures to a new ref struct called SpanOrArray, which is backed by either a ReadOnlySpan or an array. The benefit here is that we can have a ToArray method that doesn't copy if it is originally backed by an array.
This will also avoid a copy when running the ASTC decoder.
On NieR this was taking 38% of texture upload time, which it does a _lot_ of when you move between areas, so there should be a 1.6x performance boost when strictly uploading textures. No doubt this will also improve texture streaming performance in UE4 games, and maybe a small reduction with video playback.
From the numbers, it's probably possible to improve the upload rate by a further 1.6x by performing layout conversion on GPU. I'm not sure if we could improve it further than that - multithreading conversion on CPU would probably result in memory bottleneck.
This doesn't extend to buffers, since we don't convert their data on the GPU emulator side.
* Remove implicit cast to array.
* Fix some issues with CacheByRange
- Cache now clears under more circumstances, the most important being the fast path write.
- Cache supports partial clear which should help when more buffers join.
- Fixed an issue with I8->I16 conversion where it wouldn't register the buffer for use on dispose.
Should hopefully fix issues with https://github.com/Ryujinx/Ryujinx-Games-List/issues/4010 and maybe others.
* Fix collection modified exception
* Fix accidental use of parameterless constructor
* Replay DynamicState when restoring from helper shader
* Zero blend state when disabled or write mask is 0
Any difference in the blend state when blend is disabled is meaningless, but Ryujinx would compare different disabled blends and compile them as separate pipelines. This change ensures that all pipelines where blend state is meaningless record it as such, which avoids compiling a bunch of pipelines that are essentially identical.
The NVIDIA driver is pretty forgiving when it comes to silly pipeline misses like this, but other drivers don't offer the same level of kindness.
This should reduce stuttering on those drivers, and might improve overall performance very slightly due to less pipeline variants being in the hash table.
* Fix blend possibly being wrong when an attachment is unmasked
* Add Index Buffer conversion for quads to Vulkan
Also adds a reusable repeating pattern index buffer to use for non-indexed
draws, and generalizes the conversion cache for buffers.
* Fix some issues
* End render pass before conversion
* Resume transform feedback after we ensure we're in a pass.
* Always generate UInt32 type indices for topology conversion
* No it's not.
* Remove unused code
* Rely on TopologyRemap to convert quads to tris.
* Remove double newline
* Ensure render pass ends before stride or I8 conversion
* Periodically Flush Commands for Vulkan
NVIDIA's OpenGL driver has a built-in mechanism to automatically flush commands to GPU when a lot have been queued. It's also pretty inconsistent, but we'll ignore that for now.
Our Vulkan implementation only submits a command buffer (flush equivalent) when it needs to. This is typically when another command buffer needs to be sequenced after it, presenting a frame, or an edge case where we flush around GPU queries to get results sooner.
This difference in flush behaviour causes a notable difference between Vulkan and OpenGL when we have to wait for commands. In the worst case, we will wait for a sync point that has just been created. In Vulkan, this sync point is created by flushing the command buffer, and storing a waitable fence that signals its completion. Our command buffer contains _every command that we queued since the last submit_, which could be an entire frame's worth of draws.
This has a huge effect on CPU <-> GPU latency. The more commands in a command buffer, the longer we have to wait for it to complete, which results in wasted time. Because we don't know when the guest will force us to wait, we always want the smallest possible latency.
By periodically flushing, we ensure that each command buffer takes a more consistent, smaller amount of time to execute, and that the back of the GPU queue isn't as far away when we need to wait for something to happen. This also might reduce time that the GPU is left inactive while commands are being built.
The main affected game is Pokemon Sword, which got significantly faster in overworld areas due to reduced waiting time when it flushes a shadow map from the main GPU thread.
Another affected game is BOTW, which gets faster depending on the area. This game flushes textures/buffers from its game thread, which is the bottleneck.
Flush latency and throughput may be improved on other games that are inexplicably slower than OpenGL. It's possible that certain games could have their performance _decreased_ slightly due to flushes not being free, but it is unlikely.
Also, flushing to get query results sooner has been tweaked to improve the number of full draw skips that can be done. (tested in SMO)
* Remove unused variable
* Fix possible issue with early query flush
* Scale SamplesPassed counter by RT scale on report
Adds a scale factor for samples passed counter report based on the render target scale at the time. This ensures that when a game reads this counter, it appears similar to the result at 1x.
This doesn't cover cases where the the render target scale changes during the queried draws, though that might be better to handle along with other scope related issues in a future rework of counters. Games generally don't count for occlusion queries over render target changes anyways.
Fixes an issue in the Splatoon games where the special charge would scale too quickly at high res, points at the end of the game would be broken (but still provide a correct winner), and playing at a low res would make it impossible to swim in ink.
May also affect LOD scaling in The Witcher 3.
* Update Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
* Vertex Buffer Alignment part 1
* Update CacheByRange
* Add Stride Change compute shader, fix storage buffers in helpers
* An AMD exclusive
* Reword
* Change rules - stride conversion when attrs misalign
* Fix stupid mistake
* Fix background pipeline compile
* Improve a few things.
* Fix some feedback
* Address Feedback
(the shader binary didn't change when i changed the source to use the subgroup size)
* Fix bug where rewritten buffer would be disposed instantly.
This was broken by the Vulkan changes - OpenGL was building host caches at boot on one thread, which is very notably slower than when it is multithreaded.
This was caused by trying to get the program binary immediately after compilation started, which blocks. Now it does it after compilation has completed.
* SPIR-V: Initialize undefined variables with a value
Changes undefined values on spir-v shaders (caused by phi nodes) to be initialized instead of truly undefined.
Fixes an issue with NVIDIA gpus seemingly not liking when a variable is _potentially_ undefined. Not sure about the details at the moment.
Fixes:
- Tilt shift blur effect in Link's Awakening (bottom of the screen)
- Potentially block flickering on newer NVIDIA gpus in Splatoon 2? Needs testing.
Testing is welcome.
* Update Ryujinx.Graphics.Shader/CodeGen/Spirv/CodeGenContext.cs
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
* Initial commit with a lot of testing stuff.
* Partial Unmap Cleanup Part 1
* Fix some minor issues, hopefully windows tests.
* Disable partial unmap tests on macos for now
Weird issue.
* Goodbye magic number
* Add COMPlus_EnableAlternateStackCheck for tests
`COMPlus_EnableAlternateStackCheck` is needed for NullReferenceException handling to work on linux after registering the signal handler, due to how dotnet registers its own signal handler.
* Address some feedback
* Force retry when memory is mapped in memory tracking
This case existed before, but returning `false` no longer retries, so it would crash immediately after unprotecting the memory... Now, we return `true` to deliberately retry.
This case existed before (was just broken by this change) and I don't really want to look into fixing the issue right now. Technically, this means that on guest code partial unmaps will retry _due to this_ rather than hitting the handler. I don't expect this to cause any issues.
This should fix random crashes in Xenoblade Chronicles 2.
* Use IsRangeMapped
* Suppress MockMemoryManager.UnmapEvent warning
This event is not signalled by the mock memory manager.
* Remove 4kb mapping
Some games and the Mario Odyssey Multiplayer mod do this.
The SMO multiplayer mod also needs you to revert #3394 as it uses a blocking socket to receive (otherwise it hangs), and it doesn't seem to like being forced as non-blocking.
* Changes 1
* Changes 2
* Better ModifiedSequence handling
This should handle PreciseEvents properly, and simplifies a few things.
* Minor changes, remove debug log
* Handle stage.Info being null
Hopefully fixes Catherine crash
* Fix shader specialization fast texture lookup
* Fix some things.
* Address Feedback Part 1
* Make method static.
* Prefetch capabilities before spawning translation threads.
The Backend Multithreading only expects one thread to submit commands at a time. When compiling shaders, the translator may request the host GPU capabilities from the backend. It's possible for a bunch of translators to do this at the same time.
There's a caching mechanism in place so that the capabilities are only fetched once. By triggering this before spawning the thread, the async translation threads no longer try to queue onto the backend queue all at the same time.
The Capabilities do need to be checked from the GPU thread, due to OpenGL needing a context to check them, so it's not possible to call the underlying backend directly.
* Initialize the capabilities when setting the GPU thread + missing call in headless
* Remove private variables
This PR adds the alternative enum values for StencilOp. Similar to the other enums, I added these with the same names but with Gl added to the end. These are used by homebrew using Nouveau, though they might be used by games with the official Vulkan driver.
https://github.com/envytools/envytools/blob/39d90be897f41434d67277ebdf244d6bd419ecd9/rnndb/graph/nv_3ddefs.xml#L77
Fixes some broken graphics in Citra, such as missing shadows in Mario Kart 7. Likely fixes other homebrew.
If two or more threads encounter a region of memory where a read action has been registered, then they must _both_ wait on the data.
Clearing the action before it completed was causing the null check above to fail, so the action would only be run on the first thread, and the second would end up continuing without waiting. Depending on what the game does, this could be disasterous.
This fixes a regression introduced by #3302 with Pokemon Legends Arceus, and possibly Catherine. There are likely other affected games. What is fixed in that PR should still be fixed.
* Fix various issues with texture sync
A variable called _actionRegistered is used to keep track of whether a tracking action has been registered for a given texture group handle. This variable is set when the action is registered, and should be unset when it is consumed. This is used to skip registering the tracking action if it's already registered, saving some time for render targets that are modified very often.
There were two issues with this. The worst issue was that the tracking action handler exits early if the handle's modified flag is false... which means that it never reset _actionRegistered, as that was done within the Sync() method called later. The second issue was that this variable was set true after the sync action was registered, so it was technically possible for the action to run immediately, set the flag to false, then set it to true.
Both situations would lead to the action never being registered again, as the texture group handle would be sure the action is already registered. This breaks the texture for the remaining runtime, or until it is disposed.
It was also possible for a texture to register sync once, then on future frames the last modified sync number did not update. This may have caused some more minor issues.
Seems to fix the Xenoblade flashing bug. Obviously this needs a lot of testing, since it was random chance. I typically had the most luck getting it to happen by switching time of day on the event theatre screen for a while, then entering the equipment screen by pressing X on an event.
May also fix weird things like random chance air swimming in BOTW, maybe a few texture streaming bugs.
* Exchange rather than CompareExchange
OpenGL game overlays and hooks tend to make a lot of assumptions about how games present frames to the screen, since presentation in OpenGL kind of sucks and they would like to have info such as the size of the screen, or if the contents are SRGB rather than linear.
There are two ways of getting this. OBS hooks swap buffers to get a frame for video capture, but it actually checks the bound framebuffer at the time. I made sure that this matches the output framebuffer (the window) so that the output matches the size. RTSS checks the viewport size by default, but this was actually set to the last used viewport by the game, causing the OSD to fly all across the screen depending on how it was used (or res scale). The viewport is now manually set to match the output framebuffer size.
In the case of RTSS, it also loads its resources by destructively setting a pixel pack parameter without regard to what it was set to by the guest application. OpenGL state can be set for a long period of time and is not expected to be set before each call to a method, so randomly changing it isn't great practice. To fix this, I've added a line to set the pixel unpack alignment back to 4 after presentation, which should cover RTSS loading its incredibly ugly font.
- RTSS and overlays that use it should no longer cause certain textures to load incorrectly. (mario kart 8, pokemon legends arceus)
- OBS Game Capture should no longer crop the game output incorrectly, flicker randomly, or capture with incorrect gamma.
This doesn't fix issues with how RTSS reports our frame timings.
This should prevent filesystem services from blocking other services that don't have their own ServerBase. May improve filesystem related stutters in certain titles.
Improves button advanced cutscenes such as Miqol's Request in Xenoblade: DE when the game is on a network share (used to stutter when voice lines played).
Should probably be tested to make sure no mysterious bugs have been unearthed, and to see if any other filesystem related perf issues are improved.
Before, it was selecting nearest neighbour, which sounded terrible. This is likely temporary til the upsampling algorithm used by the switch is reversed.
Fixes bad audio in Skyward Sword HD.
* Add timestamp to 16-byte semaphore releases.
BOTW was reading a ulong 8 bytes after a semaphore return. Turns out this is the timestamp it was trying to do performance calculation with, so I've made it write when necessary.
This mode was also added to the DMA semaphore I added recently, as it is required by a few games. (i think quake?)
The timestamp code has been moved to GPU context. Check other games with an unusually low framerate cap or dynamic resolution to see if they have improved.
* Cast dma semaphore payload to ulong to fill the space
* Write timestamp first
Might be just worrying too much, but we don't want the applcation reading timestamp if it sees the payload before timestamp is written.
This fixes an issue where the render scale array would not be updated when technically the scales on the flat array were the same, but the start index for the vertex scales was different.
This fixes a rare bug where reporting a counter for a region containing 0 draws could deadlock the GPU. If this write overlaps with a tracking action, then the GPU could end up waiting on something that it's meant to do in the future, so it would just get stuck.
Before, this reported immediately and wrote the result to guest memory (tracked) from the backend thread.
The backend thread cannot be allowed to trigger read actions that wait on the GPU, as it will end up waiting on itself, and never advancing.
In the case of backend multithreading's `SyncMap`, it would try to wait for a backend sync object that does not yet exist, as the sync object would exist according to the GPU and tracking, but it has not yet been created by the backend (and never will be, since it's stuck).
The fix is to queue the 0 draw event just like any other, its _bufferMap value is just forced to 0.
This affects games that use Conditional Rendering: SMO, Splatoon 2, MK8. Was generally indicated by a red message in log saying that the query result timed out after 5000 tries, but not always the cause.
This fixes some regressions caused by #2971 which caused rendered 3D texture data to be lost for most slices. Fixes issues with Xenoblade 2's colour grading, probably a ton of other games.
This also removes the check from TextureCache, making it the tiniest bit smaller (any win is a win here).
* Initial test for texture sync
* WIP new texture flushing setup
* Improve rules for incompatible overlaps
Fixes a lot of issues with Unreal Engine games. Still a few minor issues (some caused by dma fast path?) Needs docs and cleanup.
* Cleanup, improvements
Improve rules for fast DMA
* Small tweak to group together flushes of overlapping handles.
* Fixes, flush overlapping texture data for ASTC and BC4/5 compressed textures.
Fixes the new Life is Strange game.
* Flush overlaps before init data, fix 3d texture size/overlap stuff
* Fix 3D Textures, faster single layer flush
Note: nosy people can no longer merge this with Vulkan. (unless they are nosy enough to implement the new backend methods)
* Remove unused method
* Minor cleanup
* More cleanup
* Use the More Fun and Hopefully No Driver Bugs method for getting compressed tex too
This one's for metro
* Address feedback, ASTC+ETC to FormatClass
* Change offset to use Span slice rather than IntPtr Add
* Fix this too
* Add support for render scale to vertex stage.
Occasionally games read off textureSize on the vertex stage to inform the fragment shader what size a texture is without querying in there. Scales were not present in the vertex shader to correct the sizes, so games were providing the raw upscaled texture size to the fragment shader, which was incorrect.
One downside is that the fragment and vertex support buffer description must be identical, so the full size scales array must be defined when used. I don't think this will have an impact though. Another is that the fragment texture count must be updated when vertex shader textures are used. I'd like to correct this so that the update is folded into the update for the scales.
Also cleans up a bunch of things, like it making no sense to call CommitRenderScale for each stage.
Fixes render scale causing a weird offset bloom in Super Mario Party and Clubhouse Games. Clubhouse Games still has a pixelated look in a number of its games due to something else it does in the shader.
* Split out support buffer update, lazy updates.
* Commit support buffer before compute dispatch
* Remove unnecessary qualifier.
* Address Feedback
Rather than calculating this for every sampler, this PR calculates if a texture can force anisotropy when its info is set, and exposes the value via a public boolean.
This should help texture/sampler heavy games when anisotropic filtering is not Auto, like UE4 ones (or so i hear?). There is another cost where samplers are created twice when anisotropic filtering is enabled, but I'm not sure how relevant this one is.
* Limit Custom Anisotropic Filtering to only fully mipmapped textures
There's a major flaw with the anisotropic filtering setting that causes @GamerzHell9137 to report graphical bugs that otherwise wouldn't be there, because he just won't set it to Auto. This should fix those issues, hopefully.
These bugs are generally because anisotropic filtering is enabled on something that it shouldn't be, such as a post process filter or some data texture. This PR maintains two host samplers when custom AF is enabled, and only uses the forced AF one when the texture is 2d and fully mipmapped (goes down to 1x1). This is because game textures are the ideal target for this filtering, and they are typically fully mipmapped, unlike things like screen render targets which usually have 1 or just a few levels.
This also only enables AF on mipmapped samplers where the filtering is bilinear or trilinear. This should be self explanatory.
This PR also allows the changing of Anisotropic Filtering at runtime, and you can immediately see the changes. All samplers are flushed from the cache if the setting changes, causing them to be recreated with the new custom AF value. This brings it in line with our resolution scale. 😌
* Expected minimum mip count for large textures rather than all, address feedback
* Use Target rather than Info.Target
* Retrigger build?
* Fix rebase
* Another workaround for NVIDIA driver 496.13 shader bug
This might work better than the other one. Give this a test to see if it fixes/doesn't fix issues with the other one.
The problem seems to be when any variable assignment happens with a negation. `temp_1 = -temp_0;` seems to trigger weird behaviour, but `temp_1 = 0.0 - temp_0;` does not. This also might to extend towards integer types?
* Update cache version
* Add disclaimer comment
* Wording
It seems that certain games (Link's Awakening, Xenoblade DE) had their fences reached already when posting framebuffers, so the signal that a frame was ready would go out _before_ the frame was enqueued, and the render loop would fail to dequeue anything and "skip" a frame.
This was resulting in their performance lowering dramatically after some loading transitions, as a frame signal would be consumed and presentation would be one frame behind.
It's possible this might have eventually caused deadlocks in these games or others, if it happened twice.
Some games (GameMaker Studio) build texture atlases out of sprites during initialization, using the 2D copy method. These copies are done from textures loaded into memory, not rendered, so they are not scaled to begin with.
I had set srcTexture in these copies to force scaling, but really it only needs to scale if the texture already exists and was scaled by rendering or something else. I just set that to false, so it doesn't change if the texture is scaled or not. This will also avoid the destination being scaled if the source wasn't. The copy can handle mismatching scales just fine.
This prevents scaling artifacts in GMS games, and maybe others (not Super Mario Maker 2, that has another issue).
Fixes a regression from #2663 where buffer flush would not happen after a resize. Specifically caused the world map in Yoshi's Crafted World to flash.
I have other planned changes to this class so this might change soon, but this regression could affect a lot so it couldn't wait.
This fixes a potential regression with the new range list changes, where the cost for creating new ones would be rather large due to creating a 1024 size array. Also reduces cost for range list inheritance by using the first existing range list as a base, rather than creating a new one then adding both lists to it.
The growth size for the RangeList is now identical to its initial size. Every 32 elements was probably a little too common - now it is 1024 for most things and 8 for the buffer modified range list.
The Unmapped and SyncMethod methods have been changed to ensure that they behave properly if the range list is set null. Cleaned up a few calls to use the null-conditional operator.
* Replace CacheResourceWrite with more general "precise" write
The goal of CacheResourceWrite was to notify GPU resources when they were modified directly, by looking up the modified address/size in a structure and calling a method on each resource. The downside of this is that each resource cache has to be queried individually, they all have to implement their own way to do this, and it can only signal to resources using the same PhysicalMemory instance.
This PR adds the ability to signal a write as "precise" on the tracking, which signals a special handler (if present) which can be used to avoid unnecessary flush actions, or maybe even more. For buffers, precise writes specifically do not flush, and instead punch a hole in the modified range list to indicate that the data on GPU has been replaced.
The downside is that precise actions must ignore the page protection bits and always signal - as they need to notify the target resource to ignore the sequence number optimization.
I had to reintroduce the sequence number increment after I2M, as removing it was causing issues in rabbids kingdom battle. However - all resources modified by I2M are notified directly to lower their sequence number, so the problem is likely that another unrelated resource is not being properly updated. Thankfully, doing this does not affect performance in the games I tested.
This should fix regressions from #2624. Test any games that were broken by that. (RF4, rabbids kingdom battle)
I've also added a sequence number increment to ThreedClass.IncrementSyncpoint, as it seems to fix buffer corruption in OpenGL homebrew. (this was a regression from removing sequence number increment from constant buffer update - another unrelated resource thing)
* Add tests.
* Add XML docs for GpuRegionHandle
* Skip UpdateProtection if only precise actions were called
This allows precise actions to skip reprotection costs.
When a texture is deleted by falling to the bottom of the AutoDeleteCache, its data is flushed to preserve any GPU writes that occurred. This ensures that the data appears in any textures recreated in the future, but didn't account for a texture that already existed with a copy dependency.
This change forces copy dependencies to complete if a texture falls out from from the AutoDeleteCache. (not removed via overlap, as that would be wasted effort)
Fixes broken lighting caused by pausing in SMO's Metro Kingdom. May fix some other issues.
Seems like this is used as an optimized way to clear memory in homebrew applications. Unfortunately, calling the software fallback method every 8 bytes was not very optimal.
The existing EmitStore is used by passing in ZR as the register to get a 0 write.
* Fast path for Inline2Memory buffer write
This PR adds a method to PhysicalMemory that attempts to write all cached resources directly, so that memory tracking can be avoided. The goal of this is both to avoid flushing buffer data, and to avoid raising the sequence number when data is written, which causes buffer and texture handles to be re-checked.
This currently only targets buffers, with a side check on textures that falls back to a tracked write if any exist within the target range. It's not expected to write textures from here - this is just a mechanism to protect us if someone does decide to do that. It's possible to add a fast path for this in future (and for ShaderCache, once that starts using tracking)
The forced read before inline2memory begins has been skipped, as the data is fully written when the transfer is completed anyways. This allows us to flush on read in emergency situations, but still write the new data over the flushed data.
Improves performance on Xenoblade 2 and DE, which was flushing buffer data on the GPU thread when trying to write compute data. May improve performance in other games that write SSBOs from compute, and update data in the same/nearby pages often.
Super Smash Bros Ultimate should probably be tested to make sure the vertex explosions haven't returned, as I think that's what this AdvanceSequence was for.
* ForceDirty before write, to make sure data does not flush over the new write
* Array based RangeList that caches Address/EndAddress
In isolation, this was more than 2x faster than the RangeList that checks using the interface. In practice I'm seeing much better results than I expected. The array is used because checking it is slightly faster than using a list, which loses time to struct copies, but I still want that data locality.
A method has been added to the list to update the cached end address, as some users of the RangeList currently modify it dynamically.
Greatly improves performance in Super Mario Odyssey, Xenoblade and any other GPU limited games.
* Address Feedback
* Lift textures in the AutoDeleteCache for all modifications.
Before, this would only apply to render targets and texture blit. Now it applies to image stores, the fast dma copy path and any other type of modification.
Image store always at least has one reference in the texture pool, so the function of the AutoDeleteCache keeping textures _alive_ is not useful, but a very important function for a while has been its use to flush textures in order of modification when they are dereferenced, so that their data is not lost.
Before, textures populated using image stores were being dereferenced and reloaded as garbage. Now, when these textures are dereferenced, their data will be put back into memory, and everything stays intact.
Fixes lighting breaking when switching levels in THPS1+2, and potentially some more UE4 games. I've tested a bunch more games for regressions and performance impact, but they all seem fine.
* Lift copy srcTexture so that it doesn't remain referenceless
* Perform lift before reference count change on unbind.
It's important to lift on unbind as that is the moment the texture was truly last modified, but definitely not after releasing every single reference.
* Fix TXQ for 3D textures.
Assumes the texture is 3D if the component mask contains Z.
This fixes a bug in UE4 games where parts of the map had garbage pointers to lighting voxels, as the lookup 3D texture was not being initialized. Most notable game is THPS1+2.
May need another PR to keep image store data alive and properly flush it in order using the AutoDeleteCache.
* Get sampler type for TextureSize from bound textures.
* Initial Implementation
* Further improvements (no support for float/64-bit types)
* Merge atomic and reduce instructions, add missing format switch
* Fix rebase issues.
* Not used.
* Whoops. Fixed.
* Partial implementation of inc/dec, cleanup and TODOs
* Remove testing path
* Address Feedback
* Avoid deleting textures when their data does not overlap.
It's possible that while two textures start and end addresses indicate an overlap, that the actual data contained within them is sparse due to a layer stride. One such possibility is array slices of a cubemap at different mip levels - they overlap on a whole, but the actual texture data fills the gaps between each other's layers rather than actually overlapping.
This fixes issues with UE4 games having incorrect lighting (solid white screen or really dark shadows). There are still remaining issues with games that use the 3D texture prebaked lighting, such as THPS1+2.
This PR also fixes a bug with TexturePool's resized texture handling where the base level in the descriptor was not considered.
* AllRegions granularity for 3d textures is now by level rather than by slice.
* Address feedback
* Only reupload the texture scale array if it changes.
Before, this would be called all the time if any shader needed a scale value. The cost of doing this has increased with threaded-gal, as the scale array is copied to a span pool, and it's was called on pretty much every draw sometimes.
This improves GPU performance in games, scaled or not. Most affected game seems to be Xenoblade Chronicles: Definitive Edition.
* Just use = instead of |=
* Initial Implementation
About as fast as nvidia GL multithreading, can be improved with faster command queuing.
* Struct based command list
Speeds up a bit. Still a lot of time lost to resource copy.
* Do shader init while the render thread is active.
* Introduce circular span pool V1
Ideally should be able to use structs instead of references for storing these spans on commands. Will try that next.
* Refactor SpanRef some more
Use a struct to represent SpanRef, rather than a reference.
* Flush buffers on background thread
* Use a span for UpdateRenderScale.
Much faster than copying the array.
* Calculate command size using reflection
* WIP parallel shaders
* Some minor optimisation
* Only 2 max refs per command now.
The command with 3 refs is gone. 😌
* Don't cast on the GPU side
* Remove redundant casts, force sync on window present
* Fix Shader Cache
* Fix host shader save.
* Fixup to work with new renderer stuff
* Make command Run static, use array of delegates as lookup
Profile says this takes less time than the previous way.
* Bring up to date
* Add settings toggle. Fix Muiltithreading Off mode.
* Fix warning.
* Release tracking lock for flushes
* Fix Conditional Render fast path with threaded gal
* Make handle iteration safe when releasing the lock
This is mostly temporary.
* Attempt to set backend threading on driver
Only really works on nvidia before launching a game.
* Fix race condition with BufferModifiedRangeList, exceptions in tracking actions
* Update buffer set commands
* Some cleanup
* Only use stutter workaround when using opengl renderer non-threaded
* Add host-conditional reservation of counter events
There has always been the possibility that conditional rendering could use a query object just as it is disposed by the counter queue. This change makes it so that when the host decides to use host conditional rendering, the query object is reserved so that it cannot be deleted. Counter events can optionally start reserved, as the threaded implementation can reserve them before the backend creates them, and there would otherwise be a short amount of time where the counter queue could dispose the event before a call to reserve it could be made.
* Address Feedback
* Make counter flush tracked again.
Hopefully does not cause any issues this time.
* Wait for FlushTo on the main queue thread.
Currently assumes only one thread will want to FlushTo (in this case, the GPU thread)
* Add SDL2 headless integration
* Add HLE macro commands.
Co-authored-by: Mary <mary@mary.zone>
This seems to be the default value when the vertex attribute is disabled, or components aren't defined.
This fixes a regression from #2307 in SMO where a plant in the Wooded Kingdom would draw slightly differently in the depth prepass, leading to depth test failing later on.
GDK has stated that the specific case in Gundam only expects x and y to be 0, and Vulkan's undefined value for z does appear to be 0 when a vertex attribute type does not have that component, hence the value (0, 0, 0, 1).
This worked in Vulkan despite also providing an all 0s attribute due to the vertex attribute binding being R32Float, so the other values were undefined. It should be changed there separately.
This greatly reduces memory usage in games that aggressively reuse memory without removing dead textures from the pool, such as the Xenoblade games, UE3 games, and to a lesser extent, UE4/unity games.
This change stops memory usage from ballooning in xenoblade and some other games. It will also reduce texture view/dependency complexity in some games - for example in MK8D it will reduce the number of surface copies between lighting cubemaps generated for actors.
There shouldn't be any performance impact from doing this, though the deletion and creation of textures could be improved by improving the OpenGL texture storage cache, which is very simple and limited right now. This will be improved in future.
Another potential error has been fixed with the texture cache, which could prevent data loss when data is interchangably written to textures from both the GPU and CPU. It was possible that the dirty flag for a texture would be consumed without the data being synchronized on next use, due to the old overlap check. This check no longer consumes the dirty flag.
Please test a bunch of games to make sure they still work, and there are no performance regressions.
Got this the wrong way round - was causing games to try synchronize mipmap levels of like 52 on a 3d texture with 6 levels. Also, corrected the variable name in the method that _was_ working.
* Use "Undesired" scale mode for certain textures rather than blacklisting
* Nit
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
* Return mapped buffer pointer directly for flush, WriteableRegion for textures
A few changes here to generally improve performance, even for platforms not using the persistent buffer flush.
- Texture and buffer flush now return a ReadOnlySpan<byte>. It's guaranteed that this span is pinned in memory, but it will be overwritten on the next flush from that thread, so it is expected that the data is used before calling again.
- As a result, persistent mappings no longer copy to a new array - rather the persistent map is returned directly as a Span<>. A similar host array is used for the glGet flushes instead of allocating new arrays each time.
- Texture flushes now do their layout conversion into a WriteableRegion when the texture is not MultiRange, which allows the flush to happen directly into guest memory rather than into a temporary span, then copied over. This avoids another copy when doing layout conversion.
Overall, this saves 1 data copy for buffer flush, 1 copy for linear textures with matching source/target stride, and 2 copies for block textures or linear textures with mismatching strides.
* Fix tests
* Fix array pointer for Mesa/Intel path
* Address some feedback
* Update method for getting array pointer.
It seems like this method of flushing data is much slower on Mesa drivers, and slightly slower on Intel Windows. Have not tested Intel Mesa, but I'm assuming it is the same as AMD.
This also adds vendor detection for AMD on Unix, which counted as "Unknown" before.
* Honour copy dependencies when switching render target
When switching from one render target to another, when both have a copy dependency to each other, a copy can be deferred on the second target when unbinding the first.
Before, this would not be honoured before binding the new texture, so the copy would stay deferred until the render targets change again, at which point it would copy in old data and essentially clear all the draws done during that time.
This change runs synchronize memory to make sure that copies are honoured. This can cause a redundant copy, but it's better than it breaking for now.
This should fix miiedit on AMD/Intel GPUs on windows. May fix other games, or perhaps rare copy dependency bugs on NVIDIA too.
* Address feedback
This greatly speeds up games that constantly resize buffers, and removes stuttering on games that resize large buffers occasionally:
- Large improvement on Super Mario 3D All-Stars (#1663 needed for best performance)
- Improvement to Hyrule Warriors: AoC, and UE4 games. These games can still stutter due to texture creation/loading.
- Small improvement to other games, potential 1-frame stutters avoided.
`ForceSynchronizeMemory`, which was added with POWER, is no longer needed. Some tests have been added for the MultiRegionHandle.
Improves OpenGL performance on FAST RMX and Xenoblade DE/2. Will probably only work on NVIDIA GPUs, but the emulated quads path will still be valid for other GPUs. Note: SLOW RMX gets a bit faster in handheld mode.
I'd recommend checking on platforms without supported host quads to make sure a GL error is actually thrown when attempting GL.Begin(PrimitiveType.Quads)
* Refactoring of KMemoryManager class
* Replace some trivial uses of DRAM address with VA
* Get rid of GetDramAddressFromVa
* Abstracting more operations on derived page table class
* Run auto-format on KPageTableBase
* Managed to make TryConvertVaToPa private, few uses remains now
* Implement guest physical pages ref counting, remove manual freeing
* Make DoMmuOperation private and call new abstract methods only from the base class
* Pass pages count rather than size on Map/UnmapMemory
* Change memory managers to take host pointers
* Fix a guest memory leak and simplify KPageTable
* Expose new methods for host range query and mapping
* Some refactoring of MapPagesFromClientProcess to allow proper page ref counting and mapping without KPageLists
* Remove more uses of AddVaRangeToPageList, now only one remains (shared memory page checking)
* Add a SharedMemoryStorage class, will be useful for host mapping
* Sayonara AddVaRangeToPageList, you served us well
* Start to implement host memory mapping (WIP)
* Support memory tracking through host exception handling
* Fix some access violations from HLE service guest memory access and CPU
* Fix memory tracking
* Fix mapping list bugs, including a race and a error adding mapping ranges
* Simple page table for memory tracking
* Simple "volatile" region handle mode
* Update UBOs directly (experimental, rough)
* Fix the overlap check
* Only set non-modified buffers as volatile
* Fix some memory tracking issues
* Fix possible race in MapBufferFromClientProcess (block list updates were not locked)
* Write uniform update to memory immediately, only defer the buffer set.
* Fix some memory tracking issues
* Pass correct pages count on shared memory unmap
* Armeilleure Signal Handler v1 + Unix changes
Unix currently behaves like windows, rather than remapping physical
* Actually check if the host platform is unix
* Fix decommit on linux.
* Implement windows 10 placeholder shared memory, fix a buffer issue.
* Make PTC version something that will never match with master
* Remove testing variable for block count
* Add reference count for memory manager, fix dispose
Can still deadlock with OpenAL
* Add address validation, use page table for mapped check, add docs
Might clean up the page table traversing routines.
* Implement batched mapping/tracking.
* Move documentation, fix tests.
* Cleanup uniform buffer update stuff.
* Remove unnecessary assignment.
* Add unsafe host mapped memory switch
On by default. Would be good to turn this off for untrusted code (homebrew, exefs mods) and give the user the option to turn it on manually, though that requires some UI work.
* Remove C# exception handlers
They have issues due to current .NET limitations, so the meilleure one fully replaces them for now.
* Fix MapPhysicalMemory on the software MemoryManager.
* Null check for GetHostAddress, docs
* Add configuration for setting memory manager mode (not in UI yet)
* Add config to UI
* Fix type mismatch on Unix signal handler code emit
* Fix 6GB DRAM mode.
The size can be greater than `uint.MaxValue` when the DRAM is >4GB.
* Address some feedback.
* More detailed error if backing memory cannot be mapped.
* SetLastError on all OS functions for consistency
* Force pages dirty with UBO update instead of setting them directly.
Seems to be much faster across a few games. Need retesting.
* Rebase, configuration rework, fix mem tracking regression
* Fix race in FreePages
* Set memory managers null after decrementing ref count
* Remove readonly keyword, as this is now modified.
* Use a local variable for the signal handler rather than a register.
* Fix bug with buffer resize, and index/uniform buffer binding.
Should fix flickering in games.
* Add InvalidAccessHandler to MemoryTracking
Doesn't do anything yet
* Call invalid access handler on unmapped read/write.
Same rules as the regular memory manager.
* Make unsafe mapped memory its own MemoryManagerType
* Move FlushUboDirty into UpdateState.
* Buffer dirty cache, rather than ubo cache
Much cleaner, may be reusable for Inline2Memory updates.
* This doesn't return anything anymore.
* Add sigaction remove methods, correct a few function signatures.
* Return empty list of physical regions for size 0.
* Also on AddressSpaceManager
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
* Compare aligned size for largest mip level when considering sampler resize
When selecting a texture that's a view for a sampler resize, we should take care that resizing it doesn't change the aligned size of any larger mip levels.
This PR covers two cases:
- When creating a view of the texture, we check that the aligned size of the view shifted up to level 0 still matches the aligned size of the container. If it does not, a copy dependency is created rather than resizing.
- When searching for a texture for sampler, textures that do _not_ match our aligned size when both are shifted up by its base level are not considered an exact match, as resizing the found texture will cause the mip 0 aligned size to change. It will create a copy dependency view instead.
Fixes graphical errors and crashes (on flush) in various Unity games that use render-to-texture.
* Move shared code to its own method.
* Pass CbufSlot when getting info from the texture descriptor
Fixes some issues with bindless textures, when CbufSlot is not equal to the current TextureBufferIndex.
Specifically fixes a random chance of full screen colour flickering in Super Mario Party.
* Apply suggestions from code review
Oops
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Co-authored-by: gdkchan <gab.dark.100@gmail.com>
riperiperiandGitHub
