Ryujinx/Ryujinx.Graphics.Gpu/Image/TextureManager.cs

1195 lines
48 KiB
C#
Raw Normal View History

2019-10-12 23:02:07 -07:00
using Ryujinx.Common;
using Ryujinx.Graphics.GAL;
using Ryujinx.Graphics.Gpu.Image;
using Ryujinx.Graphics.Gpu.Memory;
using Ryujinx.Graphics.Gpu.State;
using Ryujinx.Graphics.Texture;
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
using Ryujinx.Memory.Range;
2019-10-12 23:02:07 -07:00
using System;
namespace Ryujinx.Graphics.Gpu.Image
{
/// <summary>
/// Texture manager.
/// </summary>
2019-12-31 14:09:49 -08:00
class TextureManager : IDisposable
2019-10-12 23:02:07 -07:00
{
private struct OverlapInfo
{
public TextureViewCompatibility Compatibility { get; }
public int FirstLayer { get; }
public int FirstLevel { get; }
public OverlapInfo(TextureViewCompatibility compatibility, int firstLayer, int firstLevel)
{
Compatibility = compatibility;
FirstLayer = firstLayer;
FirstLevel = firstLevel;
}
}
private const int OverlapsBufferInitialCapacity = 10;
private const int OverlapsBufferMaxCapacity = 10000;
private readonly GpuContext _context;
2019-10-12 23:02:07 -07:00
private readonly TextureBindingsManager _cpBindingsManager;
private readonly TextureBindingsManager _gpBindingsManager;
2019-10-12 23:02:07 -07:00
private readonly Texture[] _rtColors;
2019-10-12 23:02:07 -07:00
private Texture _rtDepthStencil;
2019-10-12 23:02:07 -07:00
private readonly ITexture[] _rtHostColors;
private ITexture _rtHostDs;
private readonly RangeList<Texture> _textures;
2019-10-12 23:02:07 -07:00
private Texture[] _textureOverlaps;
private OverlapInfo[] _overlapInfo;
private readonly AutoDeleteCache _cache;
2019-10-12 23:02:07 -07:00
/// <summary>
/// The scaling factor applied to all currently bound render targets.
/// </summary>
public float RenderTargetScale { get; private set; } = 1f;
/// <summary>
/// Constructs a new instance of the texture manager.
/// </summary>
/// <param name="context">The GPU context that the texture manager belongs to</param>
public TextureManager(GpuContext context)
2019-10-12 23:02:07 -07:00
{
_context = context;
2019-10-12 23:02:07 -07:00
TexturePoolCache texturePoolCache = new TexturePoolCache(context);
_cpBindingsManager = new TextureBindingsManager(context, texturePoolCache, isCompute: true);
_gpBindingsManager = new TextureBindingsManager(context, texturePoolCache, isCompute: false);
2019-10-12 23:02:07 -07:00
_rtColors = new Texture[Constants.TotalRenderTargets];
_rtHostColors = new ITexture[Constants.TotalRenderTargets];
_textures = new RangeList<Texture>();
2019-10-12 23:02:07 -07:00
_textureOverlaps = new Texture[OverlapsBufferInitialCapacity];
_overlapInfo = new OverlapInfo[OverlapsBufferInitialCapacity];
2019-10-12 23:02:07 -07:00
_cache = new AutoDeleteCache();
}
2019-10-12 23:02:07 -07:00
/// <summary>
/// Sets texture bindings on the compute pipeline.
/// </summary>
/// <param name="bindings">The texture bindings</param>
public void SetComputeTextures(TextureBindingInfo[] bindings)
{
_cpBindingsManager.SetTextures(0, bindings);
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Sets texture bindings on the graphics pipeline.
/// </summary>
/// <param name="stage">The index of the shader stage to bind the textures</param>
/// <param name="bindings">The texture bindings</param>
public void SetGraphicsTextures(int stage, TextureBindingInfo[] bindings)
2019-10-12 23:02:07 -07:00
{
_gpBindingsManager.SetTextures(stage, bindings);
}
2019-10-12 23:02:07 -07:00
/// <summary>
/// Sets image bindings on the compute pipeline.
/// </summary>
/// <param name="bindings">The image bindings</param>
public void SetComputeImages(TextureBindingInfo[] bindings)
{
_cpBindingsManager.SetImages(0, bindings);
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Sets image bindings on the graphics pipeline.
/// </summary>
/// <param name="stage">The index of the shader stage to bind the images</param>
/// <param name="bindings">The image bindings</param>
public void SetGraphicsImages(int stage, TextureBindingInfo[] bindings)
2019-10-12 23:02:07 -07:00
{
_gpBindingsManager.SetImages(stage, bindings);
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Sets the texture constant buffer index on the compute pipeline.
/// </summary>
/// <param name="index">The texture constant buffer index</param>
public void SetComputeTextureBufferIndex(int index)
2019-10-12 23:02:07 -07:00
{
_cpBindingsManager.SetTextureBufferIndex(index);
}
2019-10-12 23:02:07 -07:00
/// <summary>
/// Sets the texture constant buffer index on the graphics pipeline.
/// </summary>
/// <param name="index">The texture constant buffer index</param>
public void SetGraphicsTextureBufferIndex(int index)
{
_gpBindingsManager.SetTextureBufferIndex(index);
}
2019-10-12 23:02:07 -07:00
/// <summary>
/// Sets the current sampler pool on the compute pipeline.
/// </summary>
/// <param name="gpuVa">The start GPU virtual address of the sampler pool</param>
/// <param name="maximumId">The maximum ID of the sampler pool</param>
2019-12-30 09:44:22 -08:00
/// <param name="samplerIndex">The indexing type of the sampler pool</param>
public void SetComputeSamplerPool(ulong gpuVa, int maximumId, SamplerIndex samplerIndex)
{
_cpBindingsManager.SetSamplerPool(gpuVa, maximumId, samplerIndex);
}
2019-10-12 23:02:07 -07:00
/// <summary>
/// Sets the current sampler pool on the graphics pipeline.
/// </summary>
/// <param name="gpuVa">The start GPU virtual address of the sampler pool</param>
/// <param name="maximumId">The maximum ID of the sampler pool</param>
2019-12-30 09:44:22 -08:00
/// <param name="samplerIndex">The indexing type of the sampler pool</param>
public void SetGraphicsSamplerPool(ulong gpuVa, int maximumId, SamplerIndex samplerIndex)
{
_gpBindingsManager.SetSamplerPool(gpuVa, maximumId, samplerIndex);
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Sets the current texture pool on the compute pipeline.
/// </summary>
/// <param name="gpuVa">The start GPU virtual address of the texture pool</param>
/// <param name="maximumId">The maximum ID of the texture pool</param>
public void SetComputeTexturePool(ulong gpuVa, int maximumId)
2019-10-12 23:02:07 -07:00
{
_cpBindingsManager.SetTexturePool(gpuVa, maximumId);
}
2019-10-12 23:02:07 -07:00
/// <summary>
/// Sets the current texture pool on the graphics pipeline.
/// </summary>
/// <param name="gpuVa">The start GPU virtual address of the texture pool</param>
/// <param name="maximumId">The maximum ID of the texture pool</param>
public void SetGraphicsTexturePool(ulong gpuVa, int maximumId)
{
_gpBindingsManager.SetTexturePool(gpuVa, maximumId);
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Sets the render target color buffer.
/// </summary>
/// <param name="index">The index of the color buffer to set (up to 8)</param>
/// <param name="color">The color buffer texture</param>
/// <returns>True if render target scale must be updated.</returns>
public bool SetRenderTargetColor(int index, Texture color)
2019-10-12 23:02:07 -07:00
{
bool hasValue = color != null;
bool changesScale = (hasValue != (_rtColors[index] != null)) || (hasValue && RenderTargetScale != color.ScaleFactor);
2019-10-12 23:02:07 -07:00
_rtColors[index] = color;
return changesScale || (hasValue && color.ScaleMode != TextureScaleMode.Blacklisted && color.ScaleFactor != GraphicsConfig.ResScale);
}
/// <summary>
/// Gets the first available bound colour target, or the depth stencil target if not present.
/// </summary>
/// <returns>The first bound colour target, otherwise the depth stencil target</returns>
public Texture GetAnyRenderTarget()
{
return _rtColors[0] ?? _rtDepthStencil;
}
/// <summary>
/// Updates the Render Target scale, given the currently bound render targets.
/// This will update scale to match the configured scale, scale textures that are eligible but not scaled,
/// and propagate blacklisted status from one texture to the ones bound with it.
/// </summary>
/// <param name="singleUse">If this is not -1, it indicates that only the given indexed target will be used.</param>
public void UpdateRenderTargetScale(int singleUse)
{
// Make sure all scales for render targets are at the highest they should be. Blacklisted targets should propagate their scale to the other targets.
bool mismatch = false;
bool blacklisted = false;
bool hasUpscaled = false;
float targetScale = GraphicsConfig.ResScale;
void ConsiderTarget(Texture target)
{
if (target == null) return;
float scale = target.ScaleFactor;
switch (target.ScaleMode)
{
case TextureScaleMode.Blacklisted:
mismatch |= scale != 1f;
blacklisted = true;
break;
case TextureScaleMode.Eligible:
mismatch = true; // We must make a decision.
break;
case TextureScaleMode.Scaled:
hasUpscaled = true;
mismatch |= scale != targetScale; // If the target scale has changed, reset the scale for all targets.
break;
}
}
if (singleUse != -1)
{
// If only one target is in use (by a clear, for example) the others do not need to be checked for mismatching scale.
ConsiderTarget(_rtColors[singleUse]);
}
else
{
foreach (Texture color in _rtColors)
{
ConsiderTarget(color);
}
}
ConsiderTarget(_rtDepthStencil);
mismatch |= blacklisted && hasUpscaled;
if (blacklisted)
{
targetScale = 1f;
}
if (mismatch)
{
if (blacklisted)
{
// Propagate the blacklisted state to the other textures.
foreach (Texture color in _rtColors)
{
color?.BlacklistScale();
}
_rtDepthStencil?.BlacklistScale();
}
else
{
// Set the scale of the other textures.
foreach (Texture color in _rtColors)
{
color?.SetScale(targetScale);
}
_rtDepthStencil?.SetScale(targetScale);
}
}
RenderTargetScale = targetScale;
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Sets the render target depth-stencil buffer.
/// </summary>
/// <param name="depthStencil">The depth-stencil buffer texture</param>
/// <returns>True if render target scale must be updated.</returns>
public bool SetRenderTargetDepthStencil(Texture depthStencil)
2019-10-12 23:02:07 -07:00
{
bool hasValue = depthStencil != null;
bool changesScale = (hasValue != (_rtDepthStencil != null)) || (hasValue && RenderTargetScale != depthStencil.ScaleFactor);
2019-10-12 23:02:07 -07:00
_rtDepthStencil = depthStencil;
return changesScale || (hasValue && depthStencil.ScaleMode != TextureScaleMode.Blacklisted && depthStencil.ScaleFactor != GraphicsConfig.ResScale);
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Commits bindings on the compute pipeline.
/// </summary>
public void CommitComputeBindings()
2019-10-12 23:02:07 -07:00
{
// Every time we switch between graphics and compute work,
// we must rebind everything.
// Since compute work happens less often, we always do that
// before and after the compute dispatch.
_cpBindingsManager.Rebind();
_cpBindingsManager.CommitBindings();
_gpBindingsManager.Rebind();
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Commits bindings on the graphics pipeline.
/// </summary>
public void CommitGraphicsBindings()
2019-10-12 23:02:07 -07:00
{
_gpBindingsManager.CommitBindings();
2019-10-12 23:02:07 -07:00
UpdateRenderTargets();
2019-10-12 23:02:07 -07:00
}
/// <summary>
/// Gets a texture descriptor used on the compute pipeline.
/// </summary>
/// <param name="state">Current GPU state</param>
/// <param name="handle">Shader "fake" handle of the texture</param>
/// <returns>The texture descriptor</returns>
public TextureDescriptor GetComputeTextureDescriptor(GpuState state, int handle)
{
return _cpBindingsManager.GetTextureDescriptor(state, 0, handle);
}
/// <summary>
/// Gets a texture descriptor used on the graphics pipeline.
/// </summary>
/// <param name="state">Current GPU state</param>
/// <param name="stageIndex">Index of the shader stage where the texture is bound</param>
/// <param name="handle">Shader "fake" handle of the texture</param>
/// <returns>The texture descriptor</returns>
public TextureDescriptor GetGraphicsTextureDescriptor(GpuState state, int stageIndex, int handle)
{
return _gpBindingsManager.GetTextureDescriptor(state, stageIndex, handle);
}
/// <summary>
/// Update host framebuffer attachments based on currently bound render target buffers.
/// </summary>
public void UpdateRenderTargets()
2019-10-12 23:02:07 -07:00
{
bool anyChanged = false;
if (_rtHostDs != _rtDepthStencil?.HostTexture)
{
_rtHostDs = _rtDepthStencil?.HostTexture;
anyChanged = true;
}
for (int index = 0; index < _rtColors.Length; index++)
2019-10-12 23:02:07 -07:00
{
ITexture hostTexture = _rtColors[index]?.HostTexture;
2019-10-12 23:02:07 -07:00
if (_rtHostColors[index] != hostTexture)
2019-10-12 23:02:07 -07:00
{
_rtHostColors[index] = hostTexture;
2019-10-12 23:02:07 -07:00
anyChanged = true;
}
}
2019-10-12 23:02:07 -07:00
if (anyChanged)
{
_context.Renderer.Pipeline.SetRenderTargets(_rtHostColors, _rtHostDs);
2019-10-12 23:02:07 -07:00
}
}
/// <summary>
/// Determines if a given texture is eligible for upscaling from its info.
/// </summary>
/// <param name="info">The texture info to check</param>
/// <returns>True if eligible</returns>
public bool IsUpscaleCompatible(TextureInfo info)
{
return (info.Target == Target.Texture2D || info.Target == Target.Texture2DArray) && !info.FormatInfo.IsCompressed && UpscaleSafeMode(info);
}
/// <summary>
/// Determines if a given texture is "safe" for upscaling from its info.
/// Note that this is different from being compatible - this elilinates targets that would have detrimental effects when scaled.
/// </summary>
/// <param name="info">The texture info to check</param>
/// <returns>True if safe</returns>
public bool UpscaleSafeMode(TextureInfo info)
{
// While upscaling works for all targets defined by IsUpscaleCompatible, we additionally blacklist targets here that
// may have undesirable results (upscaling blur textures) or simply waste GPU resources (upscaling texture atlas).
if (info.Levels > 3)
{
// Textures with more than 3 levels are likely to be game textures, rather than render textures.
// Small textures with full mips are likely to be removed by the next check.
return false;
}
if (info.Width < 8 || info.Height < 8)
{
// Discount textures with small dimensions.
return false;
}
if (!(info.FormatInfo.Format.IsDepthOrStencil() || info.FormatInfo.Components == 1))
{
// Discount square textures that aren't depth-stencil like. (excludes game textures, cubemap faces, most 3D texture LUT, texture atlas)
// Detect if the texture is possibly square. Widths may be aligned, so to remove the uncertainty we align both the width and height.
int widthAlignment = (info.IsLinear ? 32 : 64) / info.FormatInfo.BytesPerPixel;
bool possiblySquare = BitUtils.AlignUp(info.Width, widthAlignment) == BitUtils.AlignUp(info.Height, widthAlignment);
if (possiblySquare)
{
return false;
}
}
int aspect = (int)Math.Round((info.Width / (float)info.Height) * 9);
if (aspect == 16 && info.Height < 360)
{
// Targets that are roughly 16:9 can only be rescaled if they're equal to or above 360p. (excludes blur and bloom textures)
return false;
}
return true;
}
/// <summary>
/// Handles removal of textures written to a memory region being unmapped.
/// </summary>
/// <param name="sender">Sender object</param>
/// <param name="e">Event arguments</param>
public void MemoryUnmappedHandler(object sender, UnmapEventArgs e)
{
Texture[] overlaps = new Texture[10];
int overlapCount;
lock (_textures)
{
overlapCount = _textures.FindOverlaps(_context.MemoryManager.Translate(e.Address), e.Size, ref overlaps);
}
for (int i = 0; i < overlapCount; i++)
{
overlaps[i].Unmapped();
}
}
/// <summary>
/// Tries to find an existing texture, or create a new one if not found.
/// </summary>
/// <param name="copyTexture">Copy texture to find or create</param>
/// <param name="formatInfo">Format information of the copy texture</param>
/// <param name="preferScaling">Indicates if the texture should be scaled from the start</param>
/// <param name="sizeHint">A hint indicating the minimum used size for the texture</param>
/// <returns>The texture</returns>
public Texture FindOrCreateTexture(CopyTexture copyTexture, FormatInfo formatInfo, bool preferScaling = true, Size? sizeHint = null)
2019-10-12 23:02:07 -07:00
{
ulong address = _context.MemoryManager.Translate(copyTexture.Address.Pack());
GPU - Improve Memory Allocation (#1722) * Implement TreeMap from scratch. Begin implementation of MemoryBlockManager * Implement GetFreePosition using MemoryBlocks * Implementation of Memory Management using a Tree. Still some issues to work around, but promising thus far. * Resolved invalid mapping issue. Performance appears promising. * Add tick metrics * Use the logger instead * Use debug loggin instead of info. * Remove unnecessary code. Add descriptions of added functions. * Improve memory allocation even further. As well as improve speed of position fetching. * Add TreeDictionary to Ryujinx Commons Removed Unnecessary Usigns * Add a Performance Profiler + Improve ReserveFixed * Begin transition to allocation in nvdrv * Create singleton nvmemallocator * Moved Allocation into Nv Related Files As requested by gdkchan, any allocation of memory has been moved into the driver files. Mapping remains in the GPU MemoryManager. * Remove unnecessary usings * Add missing descriptions * Correct descriptions * Fix formatting. * Remove unnecessary whitespace * Formatting / Convention Updates * Changes / Fixes Made syntax and convention changes as requested by gdkchan. Fixed an issue where IsRegionUsed would return the wrong boolean. Fixed an issue where GetFreePosition was asked for an address instead of a size. * Undo commenting of Assert in shader cache * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Resolved many suggestions * Implement Improved TreeDictionary Based off of Pseudo code and custom implementations. * Rename _set to _dictionary * Remove unused code * Remove unused code. * Remove unnecessary MapLow function. * Resolve data-structure based issues * Make adjustments to memory management. Deactive de-allocation for now, it causes more harm than good. * Minor refactorings + Re-implement deallocation Also cleaned up unnecessary code. * Add Tests for TreeDictionary * Update data structure to properly balance the tree * Experimental Implementation: 1. Reduce Time to Next Node to O(1) Runtime 2. Reduce While Loop Ct To 2 (In Most Cases) * Address issues w/ Deallocating Memory * Final Build + Fully Implement Dictionary Interface for new Data Structure + Cover All Memory Allocation Edge Cases, particularly w/ Games that De-Allocate a lot. * Minor Corrections Give TreeDictionary its own count (do not depend on inner dictionary) Properly remove adjacent allocations * Add AsList * Fix bug where internal dictionary wasn't being updated w/ new node for overwritten key. * Address comments in review. * Fix issue where block wouldn't break out (Fixes UE4 issues) * Update descriptions * Update descriptions * Reduce Node visibility to protect TreeDictionary Integrity + Remove usage of struct. * Update tests to use new TreeDictionary implementation. * Remove usage of dictionary in TreeDictionary * Refactoring / Renaming * Remove unneeded memoryblock class. * Add space for while * Add space for if * Formatting / descriptions * Clarified some descriptions * Reduce visibility of memory allocator * Edit method names to make more sense as memory blocks are no longer in use. * Make names consistent. * Protect against npe when sucessorof is called against keys that don't exist. (Not in use by memory manager, this is for other prs that might use this data structure) * Possible edge-case resolve * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Update Ryujinx.HLE/HOS/Services/Nv/NvMemoryAllocator.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Reduce # of unnecessary duplicate variables / Reduce visibility of variables only internally used. * Rename count to _count * Update Description of Add method. * Fix copypasta * Address comments * Address comments * Remove whitespace * Address comments, condense variables. * Consolidate vars * Fix whitespace. * Nit * Fix exception msg * Fix arrayIndex check * Fix arrayIndex check + indexer * Remove whitespace from cast Co-authored-by: gdkchan <gab.dark.100@gmail.com>
2020-12-09 14:26:05 -08:00
if (address == MemoryManager.PteUnmapped)
2019-10-12 23:02:07 -07:00
{
return null;
}
int gobBlocksInY = copyTexture.MemoryLayout.UnpackGobBlocksInY();
int gobBlocksInZ = copyTexture.MemoryLayout.UnpackGobBlocksInZ();
int width;
if (copyTexture.LinearLayout)
{
width = copyTexture.Stride / formatInfo.BytesPerPixel;
}
else
{
width = copyTexture.Width;
}
2019-10-12 23:02:07 -07:00
TextureInfo info = new TextureInfo(
address,
width,
2019-10-12 23:02:07 -07:00
copyTexture.Height,
copyTexture.Depth,
1,
1,
1,
copyTexture.Stride,
copyTexture.LinearLayout,
gobBlocksInY,
gobBlocksInZ,
1,
Target.Texture2D,
formatInfo);
TextureSearchFlags flags = TextureSearchFlags.ForCopy;
if (preferScaling)
{
flags |= TextureSearchFlags.WithUpscale;
}
Texture texture = FindOrCreateTexture(info, flags, 0, sizeHint);
2019-10-12 23:02:07 -07:00
texture.SynchronizeMemory();
return texture;
}
/// <summary>
/// Tries to find an existing texture, or create a new one if not found.
/// </summary>
/// <param name="colorState">Color buffer texture to find or create</param>
/// <param name="samplesInX">Number of samples in the X direction, for MSAA</param>
/// <param name="samplesInY">Number of samples in the Y direction, for MSAA</param>
/// <param name="sizeHint">A hint indicating the minimum used size for the texture</param>
/// <returns>The texture</returns>
public Texture FindOrCreateTexture(RtColorState colorState, int samplesInX, int samplesInY, Size sizeHint)
2019-10-12 23:02:07 -07:00
{
ulong address = _context.MemoryManager.Translate(colorState.Address.Pack());
GPU - Improve Memory Allocation (#1722) * Implement TreeMap from scratch. Begin implementation of MemoryBlockManager * Implement GetFreePosition using MemoryBlocks * Implementation of Memory Management using a Tree. Still some issues to work around, but promising thus far. * Resolved invalid mapping issue. Performance appears promising. * Add tick metrics * Use the logger instead * Use debug loggin instead of info. * Remove unnecessary code. Add descriptions of added functions. * Improve memory allocation even further. As well as improve speed of position fetching. * Add TreeDictionary to Ryujinx Commons Removed Unnecessary Usigns * Add a Performance Profiler + Improve ReserveFixed * Begin transition to allocation in nvdrv * Create singleton nvmemallocator * Moved Allocation into Nv Related Files As requested by gdkchan, any allocation of memory has been moved into the driver files. Mapping remains in the GPU MemoryManager. * Remove unnecessary usings * Add missing descriptions * Correct descriptions * Fix formatting. * Remove unnecessary whitespace * Formatting / Convention Updates * Changes / Fixes Made syntax and convention changes as requested by gdkchan. Fixed an issue where IsRegionUsed would return the wrong boolean. Fixed an issue where GetFreePosition was asked for an address instead of a size. * Undo commenting of Assert in shader cache * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Resolved many suggestions * Implement Improved TreeDictionary Based off of Pseudo code and custom implementations. * Rename _set to _dictionary * Remove unused code * Remove unused code. * Remove unnecessary MapLow function. * Resolve data-structure based issues * Make adjustments to memory management. Deactive de-allocation for now, it causes more harm than good. * Minor refactorings + Re-implement deallocation Also cleaned up unnecessary code. * Add Tests for TreeDictionary * Update data structure to properly balance the tree * Experimental Implementation: 1. Reduce Time to Next Node to O(1) Runtime 2. Reduce While Loop Ct To 2 (In Most Cases) * Address issues w/ Deallocating Memory * Final Build + Fully Implement Dictionary Interface for new Data Structure + Cover All Memory Allocation Edge Cases, particularly w/ Games that De-Allocate a lot. * Minor Corrections Give TreeDictionary its own count (do not depend on inner dictionary) Properly remove adjacent allocations * Add AsList * Fix bug where internal dictionary wasn't being updated w/ new node for overwritten key. * Address comments in review. * Fix issue where block wouldn't break out (Fixes UE4 issues) * Update descriptions * Update descriptions * Reduce Node visibility to protect TreeDictionary Integrity + Remove usage of struct. * Update tests to use new TreeDictionary implementation. * Remove usage of dictionary in TreeDictionary * Refactoring / Renaming * Remove unneeded memoryblock class. * Add space for while * Add space for if * Formatting / descriptions * Clarified some descriptions * Reduce visibility of memory allocator * Edit method names to make more sense as memory blocks are no longer in use. * Make names consistent. * Protect against npe when sucessorof is called against keys that don't exist. (Not in use by memory manager, this is for other prs that might use this data structure) * Possible edge-case resolve * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Update Ryujinx.HLE/HOS/Services/Nv/NvMemoryAllocator.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Reduce # of unnecessary duplicate variables / Reduce visibility of variables only internally used. * Rename count to _count * Update Description of Add method. * Fix copypasta * Address comments * Address comments * Remove whitespace * Address comments, condense variables. * Consolidate vars * Fix whitespace. * Nit * Fix exception msg * Fix arrayIndex check * Fix arrayIndex check + indexer * Remove whitespace from cast Co-authored-by: gdkchan <gab.dark.100@gmail.com>
2020-12-09 14:26:05 -08:00
if (address == MemoryManager.PteUnmapped)
2019-10-12 23:02:07 -07:00
{
return null;
}
bool isLinear = colorState.MemoryLayout.UnpackIsLinear();
int gobBlocksInY = colorState.MemoryLayout.UnpackGobBlocksInY();
int gobBlocksInZ = colorState.MemoryLayout.UnpackGobBlocksInZ();
Target target;
if (colorState.MemoryLayout.UnpackIsTarget3D())
{
target = Target.Texture3D;
}
else if ((samplesInX | samplesInY) != 1)
{
target = colorState.Depth > 1
? Target.Texture2DMultisampleArray
: Target.Texture2DMultisample;
}
else
{
target = colorState.Depth > 1
? Target.Texture2DArray
: Target.Texture2D;
}
FormatInfo formatInfo = colorState.Format.Convert();
int width, stride;
// For linear textures, the width value is actually the stride.
// We can easily get the width by dividing the stride by the bpp,
// since the stride is the total number of bytes occupied by a
// line. The stride should also meet alignment constraints however,
// so the width we get here is the aligned width.
if (isLinear)
{
width = colorState.WidthOrStride / formatInfo.BytesPerPixel;
stride = colorState.WidthOrStride;
}
else
{
width = colorState.WidthOrStride;
stride = 0;
}
TextureInfo info = new TextureInfo(
address,
width,
colorState.Height,
colorState.Depth,
1,
samplesInX,
samplesInY,
stride,
isLinear,
gobBlocksInY,
gobBlocksInZ,
1,
target,
formatInfo);
int layerSize = !isLinear ? colorState.LayerSize * 4 : 0;
Texture texture = FindOrCreateTexture(info, TextureSearchFlags.WithUpscale, layerSize, sizeHint);
2019-10-12 23:02:07 -07:00
texture.SynchronizeMemory();
return texture;
}
/// <summary>
/// Tries to find an existing texture, or create a new one if not found.
/// </summary>
/// <param name="dsState">Depth-stencil buffer texture to find or create</param>
/// <param name="size">Size of the depth-stencil texture</param>
/// <param name="samplesInX">Number of samples in the X direction, for MSAA</param>
/// <param name="samplesInY">Number of samples in the Y direction, for MSAA</param>
/// <param name="sizeHint">A hint indicating the minimum used size for the texture</param>
/// <returns>The texture</returns>
public Texture FindOrCreateTexture(RtDepthStencilState dsState, Size3D size, int samplesInX, int samplesInY, Size sizeHint)
2019-10-12 23:02:07 -07:00
{
ulong address = _context.MemoryManager.Translate(dsState.Address.Pack());
GPU - Improve Memory Allocation (#1722) * Implement TreeMap from scratch. Begin implementation of MemoryBlockManager * Implement GetFreePosition using MemoryBlocks * Implementation of Memory Management using a Tree. Still some issues to work around, but promising thus far. * Resolved invalid mapping issue. Performance appears promising. * Add tick metrics * Use the logger instead * Use debug loggin instead of info. * Remove unnecessary code. Add descriptions of added functions. * Improve memory allocation even further. As well as improve speed of position fetching. * Add TreeDictionary to Ryujinx Commons Removed Unnecessary Usigns * Add a Performance Profiler + Improve ReserveFixed * Begin transition to allocation in nvdrv * Create singleton nvmemallocator * Moved Allocation into Nv Related Files As requested by gdkchan, any allocation of memory has been moved into the driver files. Mapping remains in the GPU MemoryManager. * Remove unnecessary usings * Add missing descriptions * Correct descriptions * Fix formatting. * Remove unnecessary whitespace * Formatting / Convention Updates * Changes / Fixes Made syntax and convention changes as requested by gdkchan. Fixed an issue where IsRegionUsed would return the wrong boolean. Fixed an issue where GetFreePosition was asked for an address instead of a size. * Undo commenting of Assert in shader cache * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Resolved many suggestions * Implement Improved TreeDictionary Based off of Pseudo code and custom implementations. * Rename _set to _dictionary * Remove unused code * Remove unused code. * Remove unnecessary MapLow function. * Resolve data-structure based issues * Make adjustments to memory management. Deactive de-allocation for now, it causes more harm than good. * Minor refactorings + Re-implement deallocation Also cleaned up unnecessary code. * Add Tests for TreeDictionary * Update data structure to properly balance the tree * Experimental Implementation: 1. Reduce Time to Next Node to O(1) Runtime 2. Reduce While Loop Ct To 2 (In Most Cases) * Address issues w/ Deallocating Memory * Final Build + Fully Implement Dictionary Interface for new Data Structure + Cover All Memory Allocation Edge Cases, particularly w/ Games that De-Allocate a lot. * Minor Corrections Give TreeDictionary its own count (do not depend on inner dictionary) Properly remove adjacent allocations * Add AsList * Fix bug where internal dictionary wasn't being updated w/ new node for overwritten key. * Address comments in review. * Fix issue where block wouldn't break out (Fixes UE4 issues) * Update descriptions * Update descriptions * Reduce Node visibility to protect TreeDictionary Integrity + Remove usage of struct. * Update tests to use new TreeDictionary implementation. * Remove usage of dictionary in TreeDictionary * Refactoring / Renaming * Remove unneeded memoryblock class. * Add space for while * Add space for if * Formatting / descriptions * Clarified some descriptions * Reduce visibility of memory allocator * Edit method names to make more sense as memory blocks are no longer in use. * Make names consistent. * Protect against npe when sucessorof is called against keys that don't exist. (Not in use by memory manager, this is for other prs that might use this data structure) * Possible edge-case resolve * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Update Ryujinx.HLE/HOS/Services/Nv/NvMemoryAllocator.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Reduce # of unnecessary duplicate variables / Reduce visibility of variables only internally used. * Rename count to _count * Update Description of Add method. * Fix copypasta * Address comments * Address comments * Remove whitespace * Address comments, condense variables. * Consolidate vars * Fix whitespace. * Nit * Fix exception msg * Fix arrayIndex check * Fix arrayIndex check + indexer * Remove whitespace from cast Co-authored-by: gdkchan <gab.dark.100@gmail.com>
2020-12-09 14:26:05 -08:00
if (address == MemoryManager.PteUnmapped)
2019-10-12 23:02:07 -07:00
{
return null;
}
int gobBlocksInY = dsState.MemoryLayout.UnpackGobBlocksInY();
int gobBlocksInZ = dsState.MemoryLayout.UnpackGobBlocksInZ();
Target target = (samplesInX | samplesInY) != 1
? Target.Texture2DMultisample
: Target.Texture2D;
FormatInfo formatInfo = dsState.Format.Convert();
TextureInfo info = new TextureInfo(
address,
size.Width,
size.Height,
size.Depth,
1,
samplesInX,
samplesInY,
0,
false,
gobBlocksInY,
gobBlocksInZ,
1,
target,
formatInfo);
Texture texture = FindOrCreateTexture(info, TextureSearchFlags.WithUpscale, dsState.LayerSize * 4, sizeHint);
2019-10-12 23:02:07 -07:00
texture.SynchronizeMemory();
return texture;
}
/// <summary>
/// Tries to find an existing texture, or create a new one if not found.
/// </summary>
/// <param name="info">Texture information of the texture to be found or created</param>
/// <param name="flags">The texture search flags, defines texture comparison rules</param>
/// <param name="layerSize">Size in bytes of a single texture layer</param>
/// <param name="sizeHint">A hint indicating the minimum used size for the texture</param>
/// <returns>The texture</returns>
public Texture FindOrCreateTexture(TextureInfo info, TextureSearchFlags flags = TextureSearchFlags.None, int layerSize = 0, Size? sizeHint = null)
2019-10-12 23:02:07 -07:00
{
bool isSamplerTexture = (flags & TextureSearchFlags.ForSampler) != 0;
2019-10-12 23:02:07 -07:00
bool isScalable = IsUpscaleCompatible(info);
TextureScaleMode scaleMode = TextureScaleMode.Blacklisted;
if (isScalable)
{
scaleMode = (flags & TextureSearchFlags.WithUpscale) != 0 ? TextureScaleMode.Scaled : TextureScaleMode.Eligible;
}
int sameAddressOverlapsCount;
lock (_textures)
{
// Try to find a perfect texture match, with the same address and parameters.
sameAddressOverlapsCount = _textures.FindOverlaps(info.Address, ref _textureOverlaps);
}
2019-10-12 23:02:07 -07:00
Texture texture = null;
TextureMatchQuality bestQuality = TextureMatchQuality.NoMatch;
for (int index = 0; index < sameAddressOverlapsCount; index++)
2019-10-12 23:02:07 -07:00
{
Texture overlap = _textureOverlaps[index];
TextureMatchQuality matchQuality = overlap.IsExactMatch(info, flags);
if (matchQuality == TextureMatchQuality.Perfect)
2019-10-12 23:02:07 -07:00
{
texture = overlap;
break;
}
else if (matchQuality > bestQuality)
{
texture = overlap;
bestQuality = matchQuality;
}
}
if (texture != null)
{
if (!isSamplerTexture)
{
// If not a sampler texture, it is managed by the auto delete
// cache, ensure that it is on the "top" of the list to avoid
// deletion.
_cache.Lift(texture);
}
2019-10-12 23:02:07 -07:00
ChangeSizeIfNeeded(info, texture, isSamplerTexture, sizeHint);
texture.SynchronizeMemory();
return texture;
2019-10-12 23:02:07 -07:00
}
// Calculate texture sizes, used to find all overlapping textures.
SizeInfo sizeInfo = info.CalculateSizeInfo(layerSize);
2019-10-12 23:02:07 -07:00
// Find view compatible matches.
ulong size = (ulong)sizeInfo.TotalSize;
int overlapsCount;
2019-10-12 23:02:07 -07:00
lock (_textures)
{
overlapsCount = _textures.FindOverlaps(info.Address, size, ref _textureOverlaps);
}
2019-10-12 23:02:07 -07:00
for (int index = 0; index < overlapsCount; index++)
2019-10-12 23:02:07 -07:00
{
Texture overlap = _textureOverlaps[index];
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
TextureViewCompatibility overlapCompatibility = overlap.IsViewCompatible(info, size, out int firstLayer, out int firstLevel);
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
if (overlapCompatibility == TextureViewCompatibility.Full)
2019-10-12 23:02:07 -07:00
{
TextureInfo oInfo = AdjustSizes(overlap, info, firstLevel);
2019-10-12 23:02:07 -07:00
if (!isSamplerTexture)
{
info = oInfo;
2019-10-12 23:02:07 -07:00
}
texture = overlap.CreateView(oInfo, sizeInfo, firstLayer, firstLevel);
2019-10-12 23:02:07 -07:00
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
if (overlap.IsModified)
{
texture.SignalModified();
}
ChangeSizeIfNeeded(info, texture, isSamplerTexture, sizeHint);
2019-10-12 23:02:07 -07:00
break;
}
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
else if (overlapCompatibility == TextureViewCompatibility.CopyOnly)
{
// TODO: Copy rules for targets created after the container texture. See below.
overlap.DisableMemoryTracking();
2019-10-12 23:02:07 -07:00
}
}
// No match, create a new texture.
if (texture == null)
{
texture = new Texture(_context, info, sizeInfo, scaleMode);
2019-10-12 23:02:07 -07:00
// Step 1: Find textures that are view compatible with the new texture.
// Any textures that are incompatible will contain garbage data, so they should be removed where possible.
int viewCompatible = 0;
bool setData = isSamplerTexture || overlapsCount == 0 || flags.HasFlag(TextureSearchFlags.ForCopy);
2019-10-12 23:02:07 -07:00
for (int index = 0; index < overlapsCount; index++)
2019-10-12 23:02:07 -07:00
{
Texture overlap = _textureOverlaps[index];
bool overlapInCache = overlap.CacheNode != null;
TextureViewCompatibility compatibility = texture.IsViewCompatible(overlap.Info, overlap.Size, out int firstLayer, out int firstLevel);
2019-10-12 23:02:07 -07:00
if (compatibility != TextureViewCompatibility.Incompatible)
{
if (_overlapInfo.Length != _textureOverlaps.Length)
{
Array.Resize(ref _overlapInfo, _textureOverlaps.Length);
}
_overlapInfo[viewCompatible] = new OverlapInfo(compatibility, firstLayer, firstLevel);
_textureOverlaps[viewCompatible++] = overlap;
}
else if (overlapInCache || !setData)
{
if (info.GobBlocksInZ > 1 && info.GobBlocksInZ == overlap.Info.GobBlocksInZ)
{
// Allow overlapping slices of 3D textures. Could be improved in future by making sure the textures don't overlap.
continue;
}
// The overlap texture is going to contain garbage data after we draw, or is generally incompatible.
// If the texture cannot be entirely contained in the new address space, and one of its view children is compatible with us,
// it must be flushed before removal, so that the data is not lost.
// If the texture was modified since its last use, then that data is probably meant to go into this texture.
// If the data has been modified by the CPU, then it also shouldn't be flushed.
bool modified = overlap.ConsumeModified();
2019-10-12 23:02:07 -07:00
bool flush = overlapInCache && !modified && (overlap.Address < texture.Address || overlap.EndAddress > texture.EndAddress) && overlap.HasViewCompatibleChild(texture);
2019-10-12 23:02:07 -07:00
setData |= modified || flush;
if (overlapInCache)
{
_cache.Remove(overlap, flush);
}
}
}
// We need to synchronize before copying the old view data to the texture,
// otherwise the copied data would be overwritten by a future synchronization.
texture.InitializeData(false, setData);
for (int index = 0; index < viewCompatible; index++)
{
Texture overlap = _textureOverlaps[index];
OverlapInfo oInfo = _overlapInfo[index];
if (oInfo.Compatibility != TextureViewCompatibility.Full)
{
continue; // Copy only compatibilty.
}
TextureInfo overlapInfo = AdjustSizes(texture, overlap.Info, oInfo.FirstLevel);
TextureCreateInfo createInfo = GetCreateInfo(overlapInfo, _context.Capabilities, overlap.ScaleFactor);
if (texture.ScaleFactor != overlap.ScaleFactor)
{
// A bit tricky, our new texture may need to contain an existing texture that is upscaled, but isn't itself.
// In that case, we prefer the higher scale only if our format is render-target-like, otherwise we scale the view down before copy.
texture.PropagateScale(overlap);
2019-10-12 23:02:07 -07:00
}
ITexture newView = texture.HostTexture.CreateView(createInfo, oInfo.FirstLayer, oInfo.FirstLevel);
overlap.HostTexture.CopyTo(newView, 0, 0);
// Inherit modification from overlapping texture, do that before replacing
// the view since the replacement operation removes it from the list.
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
if (overlap.IsModified)
{
texture.SignalModified();
}
overlap.ReplaceView(texture, overlapInfo, newView, oInfo.FirstLayer, oInfo.FirstLevel);
2019-10-12 23:02:07 -07:00
}
// If the texture is a 3D texture, we need to additionally copy any slice
// of the 3D texture to the newly created 3D texture.
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
if (info.Target == Target.Texture3D && viewCompatible > 0)
{
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
// TODO: This copy can currently only happen when the 3D texture is created.
// If a game clears and redraws the slices, we won't be able to copy the new data to the 3D texture.
// Disable tracking to try keep at least the original data in there for as long as possible.
texture.DisableMemoryTracking();
for (int index = 0; index < viewCompatible; index++)
{
Texture overlap = _textureOverlaps[index];
OverlapInfo oInfo = _overlapInfo[index];
if (oInfo.Compatibility != TextureViewCompatibility.Incompatible)
{
overlap.BlacklistScale();
overlap.HostTexture.CopyTo(texture.HostTexture, oInfo.FirstLayer, oInfo.FirstLevel);
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
if (overlap.IsModified)
{
texture.SignalModified();
}
}
}
}
2019-10-12 23:02:07 -07:00
}
// Sampler textures are managed by the texture pool, all other textures
// are managed by the auto delete cache.
if (!isSamplerTexture)
{
_cache.Add(texture);
}
lock (_textures)
{
_textures.Add(texture);
}
2019-10-12 23:02:07 -07:00
ShrinkOverlapsBufferIfNeeded();
2019-10-12 23:02:07 -07:00
return texture;
}
/// <summary>
/// Changes a texture's size to match the desired size for samplers,
/// or increases a texture's size to fit the region indicated by a size hint.
/// </summary>
/// <param name="info">The desired texture info</param>
/// <param name="texture">The texture to resize</param>
/// <param name="isSamplerTexture">True if the texture will be used for a sampler, false otherwise</param>
/// <param name="sizeHint">A hint indicating the minimum used size for the texture</param>
private void ChangeSizeIfNeeded(TextureInfo info, Texture texture, bool isSamplerTexture, Size? sizeHint)
{
if (isSamplerTexture)
{
// If this is used for sampling, the size must match,
// otherwise the shader would sample garbage data.
// To fix that, we create a new texture with the correct
// size, and copy the data from the old one to the new one.
if (!TextureCompatibility.SizeMatches(texture.Info, info))
{
texture.ChangeSize(info.Width, info.Height, info.DepthOrLayers);
}
}
else if (sizeHint != null)
{
// A size hint indicates that data will be used within that range, at least.
// If the texture is smaller than the size hint, it must be enlarged to meet it.
// The maximum size is provided by the requested info, which generally has an aligned size.
int width = Math.Max(texture.Info.Width, Math.Min(sizeHint.Value.Width, info.Width));
int height = Math.Max(texture.Info.Height, Math.Min(sizeHint.Value.Height, info.Height));
if (texture.Info.Width != width || texture.Info.Height != height)
{
texture.ChangeSize(width, height, info.DepthOrLayers);
}
}
}
/// <summary>
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
/// Tries to find an existing texture matching the given buffer copy destination. If none is found, returns null.
/// </summary>
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
/// <param name="tex">The texture information</param>
/// <param name="cbp">The copy buffer parameters</param>
/// <param name="swizzle">The copy buffer swizzle</param>
/// <param name="linear">True if the texture has a linear layout, false otherwise</param>
/// <returns>A matching texture, or null if there is no match</returns>
public Texture FindTexture(CopyBufferTexture tex, CopyBufferParams cbp, CopyBufferSwizzle swizzle, bool linear)
{
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
ulong address = _context.MemoryManager.Translate(cbp.DstAddress.Pack());
GPU - Improve Memory Allocation (#1722) * Implement TreeMap from scratch. Begin implementation of MemoryBlockManager * Implement GetFreePosition using MemoryBlocks * Implementation of Memory Management using a Tree. Still some issues to work around, but promising thus far. * Resolved invalid mapping issue. Performance appears promising. * Add tick metrics * Use the logger instead * Use debug loggin instead of info. * Remove unnecessary code. Add descriptions of added functions. * Improve memory allocation even further. As well as improve speed of position fetching. * Add TreeDictionary to Ryujinx Commons Removed Unnecessary Usigns * Add a Performance Profiler + Improve ReserveFixed * Begin transition to allocation in nvdrv * Create singleton nvmemallocator * Moved Allocation into Nv Related Files As requested by gdkchan, any allocation of memory has been moved into the driver files. Mapping remains in the GPU MemoryManager. * Remove unnecessary usings * Add missing descriptions * Correct descriptions * Fix formatting. * Remove unnecessary whitespace * Formatting / Convention Updates * Changes / Fixes Made syntax and convention changes as requested by gdkchan. Fixed an issue where IsRegionUsed would return the wrong boolean. Fixed an issue where GetFreePosition was asked for an address instead of a size. * Undo commenting of Assert in shader cache * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Resolved many suggestions * Implement Improved TreeDictionary Based off of Pseudo code and custom implementations. * Rename _set to _dictionary * Remove unused code * Remove unused code. * Remove unnecessary MapLow function. * Resolve data-structure based issues * Make adjustments to memory management. Deactive de-allocation for now, it causes more harm than good. * Minor refactorings + Re-implement deallocation Also cleaned up unnecessary code. * Add Tests for TreeDictionary * Update data structure to properly balance the tree * Experimental Implementation: 1. Reduce Time to Next Node to O(1) Runtime 2. Reduce While Loop Ct To 2 (In Most Cases) * Address issues w/ Deallocating Memory * Final Build + Fully Implement Dictionary Interface for new Data Structure + Cover All Memory Allocation Edge Cases, particularly w/ Games that De-Allocate a lot. * Minor Corrections Give TreeDictionary its own count (do not depend on inner dictionary) Properly remove adjacent allocations * Add AsList * Fix bug where internal dictionary wasn't being updated w/ new node for overwritten key. * Address comments in review. * Fix issue where block wouldn't break out (Fixes UE4 issues) * Update descriptions * Update descriptions * Reduce Node visibility to protect TreeDictionary Integrity + Remove usage of struct. * Update tests to use new TreeDictionary implementation. * Remove usage of dictionary in TreeDictionary * Refactoring / Renaming * Remove unneeded memoryblock class. * Add space for while * Add space for if * Formatting / descriptions * Clarified some descriptions * Reduce visibility of memory allocator * Edit method names to make more sense as memory blocks are no longer in use. * Make names consistent. * Protect against npe when sucessorof is called against keys that don't exist. (Not in use by memory manager, this is for other prs that might use this data structure) * Possible edge-case resolve * Update Ryujinx.Common/Collections/TreeDictionary.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Update Ryujinx.HLE/HOS/Services/Nv/NvMemoryAllocator.cs Co-authored-by: gdkchan <gab.dark.100@gmail.com> * Reduce # of unnecessary duplicate variables / Reduce visibility of variables only internally used. * Rename count to _count * Update Description of Add method. * Fix copypasta * Address comments * Address comments * Remove whitespace * Address comments, condense variables. * Consolidate vars * Fix whitespace. * Nit * Fix exception msg * Fix arrayIndex check * Fix arrayIndex check + indexer * Remove whitespace from cast Co-authored-by: gdkchan <gab.dark.100@gmail.com>
2020-12-09 14:26:05 -08:00
if (address == MemoryManager.PteUnmapped)
{
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
return null;
}
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
int bpp = swizzle.UnpackDstComponentsCount() * swizzle.UnpackComponentSize();
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
int addressMatches = _textures.FindOverlaps(address, ref _textureOverlaps);
for (int i = 0; i < addressMatches; i++)
{
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
Texture texture = _textureOverlaps[i];
FormatInfo format = texture.Info.FormatInfo;
if (texture.Info.DepthOrLayers > 1)
{
continue;
}
bool match;
if (linear)
{
// Size is not available for linear textures. Use the stride and end of the copy region instead.
match = texture.Info.IsLinear && texture.Info.Stride == cbp.DstStride && tex.RegionY + cbp.YCount <= texture.Info.Height;
}
else
{
// Bpp may be a mismatch between the target texture and the param.
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
// Due to the way linear strided and block layouts work, widths can be multiplied by Bpp for comparison.
// Note: tex.Width is the aligned texture size. Prefer param.XCount, as the destination should be a texture with that exact size.
bool sizeMatch = cbp.XCount * bpp == texture.Info.Width * format.BytesPerPixel && tex.Height == texture.Info.Height;
bool formatMatch = !texture.Info.IsLinear &&
texture.Info.GobBlocksInY == tex.MemoryLayout.UnpackGobBlocksInY() &&
texture.Info.GobBlocksInZ == tex.MemoryLayout.UnpackGobBlocksInZ();
match = sizeMatch && formatMatch;
}
if (match)
{
return texture;
}
}
Memory Read/Write Tracking using Region Handles (#1272) * WIP Range Tracking - Texture invalidation seems to have large problems - Buffer/Pool invalidation may have problems - Mirror memory tracking puts an additional `add` in compiled code, we likely just want to make HLE access slower if this is the final solution. - Native project is in the messiest possible location. - [HACK] JIT memory access always uses native "fast" path - [HACK] Trying some things with texture invalidation and views. It works :) Still a few hacks, messy things, slow things More work in progress stuff (also move to memory project) Quite a bit faster now. - Unmapping GPU VA and CPU VA will now correctly update write tracking regions, and invalidate textures for the former. - The Virtual range list is now non-overlapping like the physical one. - Fixed some bugs where regions could leak. - Introduced a weird bug that I still need to track down (consistent invalid buffer in MK8 ribbon road) Move some stuff. I think we'll eventually just put the dll and so for this in a nuget package. Fix rebase. [WIP] MultiRegionHandle variable size ranges - Avoid reprotecting regions that change often (needs some tweaking) - There's still a bug in buffers, somehow. - Might want different api for minimum granularity Fix rebase issue Commit everything needed for software only tracking. Remove native components. Remove more native stuff. Cleanup Use a separate window for the background context, update opentk. (fixes linux) Some experimental changes Should get things working up to scratch - still need to try some things with flush/modification and res scale. Include address with the region action. Initial work to make range tracking work Still a ton of bugs Fix some issues with the new stuff. * Fix texture flush instability There's still some weird behaviour, but it's much improved without this. (textures with cpu modified data were flushing over it) * Find the destination texture for Buffer->Texture full copy Greatly improves performance for nvdec videos (with range tracking) * Further improve texture tracking * Disable Memory Tracking for view parents This is a temporary approach to better match behaviour on master (where invalidations would be soaked up by views, rather than trigger twice) The assumption is that when views are created to a texture, they will cover all of its data anyways. Of course, this can easily be improved in future. * Introduce some tracking tests. WIP * Complete base tests. * Add more tests for multiregion, fix existing test. * Cleanup Part 1 * Remove unnecessary code from memory tracking * Fix some inconsistencies with 3D texture rule. * Add dispose tests. * Use a background thread for the background context. Rather than setting and unsetting a context as current, doing the work on a dedicated thread with signals seems to be a bit faster. Also nerf the multithreading test a bit. * Copy to texture with matching alignment This extends the copy to work for some videos with unusual size, such as tutorial videos in SMO. It will only occur if the destination texture already exists at XCount size. * Track reads for buffer copies. Synchronize new buffers before copying overlaps. * Remove old texture flushing mechanisms. Range tracking all the way, baby. * Wake the background thread when disposing. Avoids a deadlock when games are closed. * Address Feedback 1 * Separate TextureCopy instance for background thread Also `BackgroundContextWorker.InBackground` for a more sensible idenfifier for if we're in a background thread. * Add missing XML docs. * Address Feedback * Maybe I should start drinking coffee. * Some more feedback. * Remove flush warning, Refocus window after making background context
2020-10-16 13:18:35 -07:00
return null;
}
/// <summary>
/// Resizes the temporary buffer used for range list intersection results, if it has grown too much.
/// </summary>
private void ShrinkOverlapsBufferIfNeeded()
{
if (_textureOverlaps.Length > OverlapsBufferMaxCapacity)
{
Array.Resize(ref _textureOverlaps, OverlapsBufferMaxCapacity);
}
}
/// <summary>
/// Adjusts the size of the texture information for a given mipmap level,
/// based on the size of a parent texture.
/// </summary>
/// <param name="parent">The parent texture</param>
/// <param name="info">The texture information to be adjusted</param>
/// <param name="firstLevel">The first level of the texture view</param>
/// <returns>The adjusted texture information with the new size</returns>
2019-10-12 23:02:07 -07:00
private static TextureInfo AdjustSizes(Texture parent, TextureInfo info, int firstLevel)
{
// When the texture is used as view of another texture, we must
// ensure that the sizes are valid, otherwise data uploads would fail
// (and the size wouldn't match the real size used on the host API).
// Given a parent texture from where the view is created, we have the
// following rules:
// - The view size must be equal to the parent size, divided by (2 ^ l),
// where l is the first mipmap level of the view. The division result must
// be rounded down, and the result must be clamped to 1.
// - If the parent format is compressed, and the view format isn't, the
// view size is calculated as above, but the width and height of the
// view must be also divided by the compressed format block width and height.
// - If the parent format is not compressed, and the view is, the view
// size is calculated as described on the first point, but the width and height
// of the view must be also multiplied by the block width and height.
int width = Math.Max(1, parent.Info.Width >> firstLevel);
int height = Math.Max(1, parent.Info.Height >> firstLevel);
if (parent.Info.FormatInfo.IsCompressed && !info.FormatInfo.IsCompressed)
{
width = BitUtils.DivRoundUp(width, parent.Info.FormatInfo.BlockWidth);
height = BitUtils.DivRoundUp(height, parent.Info.FormatInfo.BlockHeight);
}
else if (!parent.Info.FormatInfo.IsCompressed && info.FormatInfo.IsCompressed)
{
width *= info.FormatInfo.BlockWidth;
height *= info.FormatInfo.BlockHeight;
}
int depthOrLayers;
if (info.Target == Target.Texture3D)
{
depthOrLayers = Math.Max(1, parent.Info.DepthOrLayers >> firstLevel);
}
else
{
depthOrLayers = info.DepthOrLayers;
}
return new TextureInfo(
info.Address,
width,
height,
depthOrLayers,
info.Levels,
info.SamplesInX,
info.SamplesInY,
info.Stride,
info.IsLinear,
info.GobBlocksInY,
info.GobBlocksInZ,
info.GobBlocksInTileX,
info.Target,
info.FormatInfo,
info.DepthStencilMode,
info.SwizzleR,
info.SwizzleG,
info.SwizzleB,
info.SwizzleA);
}
/// <summary>
/// Gets a texture creation information from texture information.
/// This can be used to create new host textures.
/// </summary>
/// <param name="info">Texture information</param>
/// <param name="caps">GPU capabilities</param>
/// <param name="scale">Texture scale factor, to be applied to the texture size</param>
/// <returns>The texture creation information</returns>
public static TextureCreateInfo GetCreateInfo(TextureInfo info, Capabilities caps, float scale)
2019-10-12 23:02:07 -07:00
{
FormatInfo formatInfo = TextureCompatibility.ToHostCompatibleFormat(info, caps);
2019-10-12 23:02:07 -07:00
if (info.Target == Target.TextureBuffer)
{
// We assume that the host does not support signed normalized format
// (as is the case with OpenGL), so we just use a unsigned format.
// The shader will need the appropriate conversion code to compensate.
switch (formatInfo.Format)
{
case Format.R8Snorm:
formatInfo = new FormatInfo(Format.R8Sint, 1, 1, 1, 1);
break;
case Format.R16Snorm:
formatInfo = new FormatInfo(Format.R16Sint, 1, 1, 2, 1);
break;
case Format.R8G8Snorm:
formatInfo = new FormatInfo(Format.R8G8Sint, 1, 1, 2, 2);
break;
case Format.R16G16Snorm:
formatInfo = new FormatInfo(Format.R16G16Sint, 1, 1, 4, 2);
break;
case Format.R8G8B8A8Snorm:
formatInfo = new FormatInfo(Format.R8G8B8A8Sint, 1, 1, 4, 4);
break;
case Format.R16G16B16A16Snorm:
formatInfo = new FormatInfo(Format.R16G16B16A16Sint, 1, 1, 8, 4);
break;
}
}
2019-10-12 23:02:07 -07:00
int width = info.Width / info.SamplesInX;
int height = info.Height / info.SamplesInY;
int depth = info.GetDepth() * info.GetLayers();
if (scale != 1f)
{
width = (int)MathF.Ceiling(width * scale);
height = (int)MathF.Ceiling(height * scale);
}
2019-10-12 23:02:07 -07:00
return new TextureCreateInfo(
width,
height,
depth,
info.Levels,
info.Samples,
formatInfo.BlockWidth,
formatInfo.BlockHeight,
formatInfo.BytesPerPixel,
formatInfo.Format,
info.DepthStencilMode,
info.Target,
info.SwizzleR,
info.SwizzleG,
info.SwizzleB,
info.SwizzleA);
}
/// <summary>
/// Removes a texture from the cache.
2020-01-01 07:39:09 -08:00
/// </summary>
/// <remarks>
/// This only removes the texture from the internal list, not from the auto-deletion cache.
/// It may still have live references after the removal.
2020-01-01 07:39:09 -08:00
/// </remarks>
/// <param name="texture">The texture to be removed</param>
2019-10-12 23:02:07 -07:00
public void RemoveTextureFromCache(Texture texture)
{
lock (_textures)
{
_textures.Remove(texture);
}
2019-10-12 23:02:07 -07:00
}
2019-12-31 14:09:49 -08:00
/// <summary>
/// Disposes all textures in the cache.
/// It's an error to use the texture manager after disposal.
/// </summary>
public void Dispose()
{
lock (_textures)
2019-12-31 14:09:49 -08:00
{
foreach (Texture texture in _textures)
{
texture.Dispose();
}
2019-12-31 14:09:49 -08:00
}
}
2019-10-12 23:02:07 -07:00
}
}