ryujinx/Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

1080 lines
48 KiB
C#
Raw Normal View History

using Ryujinx.Common;
using Ryujinx.Common.Logging;
using Ryujinx.Graphics.GAL;
using Ryujinx.Graphics.Gpu.Engine.Threed;
using Ryujinx.Graphics.Gpu.Memory;
using Ryujinx.Graphics.Gpu.Shader.Cache;
using Ryujinx.Graphics.Gpu.Shader.Cache.Definition;
using Ryujinx.Graphics.Shader;
using Ryujinx.Graphics.Shader.Translation;
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Runtime.InteropServices;
using System.Threading;
using System.Threading.Tasks;
namespace Ryujinx.Graphics.Gpu.Shader
{
/// <summary>
/// Memory cache of shader code.
/// </summary>
2019-12-31 23:09:49 +01:00
class ShaderCache : IDisposable
{
private const TranslationFlags DefaultFlags = TranslationFlags.DebugMode;
private readonly GpuContext _context;
private readonly ShaderDumper _dumper;
private readonly Dictionary<ulong, List<ShaderBundle>> _cpPrograms;
private readonly Dictionary<ShaderAddresses, List<ShaderBundle>> _gpPrograms;
private CacheManager _cacheManager;
private Dictionary<Hash128, ShaderBundle> _gpProgramsDiskCache;
private Dictionary<Hash128, ShaderBundle> _cpProgramsDiskCache;
Add a Multithreading layer for the GAL, multi-thread shader compilation at runtime (#2501) * Initial Implementation About as fast as nvidia GL multithreading, can be improved with faster command queuing. * Struct based command list Speeds up a bit. Still a lot of time lost to resource copy. * Do shader init while the render thread is active. * Introduce circular span pool V1 Ideally should be able to use structs instead of references for storing these spans on commands. Will try that next. * Refactor SpanRef some more Use a struct to represent SpanRef, rather than a reference. * Flush buffers on background thread * Use a span for UpdateRenderScale. Much faster than copying the array. * Calculate command size using reflection * WIP parallel shaders * Some minor optimisation * Only 2 max refs per command now. The command with 3 refs is gone. :relieved: * Don't cast on the GPU side * Remove redundant casts, force sync on window present * Fix Shader Cache * Fix host shader save. * Fixup to work with new renderer stuff * Make command Run static, use array of delegates as lookup Profile says this takes less time than the previous way. * Bring up to date * Add settings toggle. Fix Muiltithreading Off mode. * Fix warning. * Release tracking lock for flushes * Fix Conditional Render fast path with threaded gal * Make handle iteration safe when releasing the lock This is mostly temporary. * Attempt to set backend threading on driver Only really works on nvidia before launching a game. * Fix race condition with BufferModifiedRangeList, exceptions in tracking actions * Update buffer set commands * Some cleanup * Only use stutter workaround when using opengl renderer non-threaded * Add host-conditional reservation of counter events There has always been the possibility that conditional rendering could use a query object just as it is disposed by the counter queue. This change makes it so that when the host decides to use host conditional rendering, the query object is reserved so that it cannot be deleted. Counter events can optionally start reserved, as the threaded implementation can reserve them before the backend creates them, and there would otherwise be a short amount of time where the counter queue could dispose the event before a call to reserve it could be made. * Address Feedback * Make counter flush tracked again. Hopefully does not cause any issues this time. * Wait for FlushTo on the main queue thread. Currently assumes only one thread will want to FlushTo (in this case, the GPU thread) * Add SDL2 headless integration * Add HLE macro commands. Co-authored-by: Mary <mary@mary.zone>
2021-08-27 00:31:29 +02:00
private Queue<(IProgram, Action<byte[]>)> _programsToSaveQueue;
/// <summary>
/// Version of the codegen (to be changed when codegen or guest format change).
/// </summary>
private const ulong ShaderCodeGenVersion = 3106;
// Progress reporting helpers
private volatile int _shaderCount;
private volatile int _totalShaderCount;
public event Action<ShaderCacheState, int, int> ShaderCacheStateChanged;
/// <summary>
/// Creates a new instance of the shader cache.
/// </summary>
/// <param name="context">GPU context that the shader cache belongs to</param>
public ShaderCache(GpuContext context)
{
_context = context;
_dumper = new ShaderDumper();
_cpPrograms = new Dictionary<ulong, List<ShaderBundle>>();
_gpPrograms = new Dictionary<ShaderAddresses, List<ShaderBundle>>();
_gpProgramsDiskCache = new Dictionary<Hash128, ShaderBundle>();
_cpProgramsDiskCache = new Dictionary<Hash128, ShaderBundle>();
Add a Multithreading layer for the GAL, multi-thread shader compilation at runtime (#2501) * Initial Implementation About as fast as nvidia GL multithreading, can be improved with faster command queuing. * Struct based command list Speeds up a bit. Still a lot of time lost to resource copy. * Do shader init while the render thread is active. * Introduce circular span pool V1 Ideally should be able to use structs instead of references for storing these spans on commands. Will try that next. * Refactor SpanRef some more Use a struct to represent SpanRef, rather than a reference. * Flush buffers on background thread * Use a span for UpdateRenderScale. Much faster than copying the array. * Calculate command size using reflection * WIP parallel shaders * Some minor optimisation * Only 2 max refs per command now. The command with 3 refs is gone. :relieved: * Don't cast on the GPU side * Remove redundant casts, force sync on window present * Fix Shader Cache * Fix host shader save. * Fixup to work with new renderer stuff * Make command Run static, use array of delegates as lookup Profile says this takes less time than the previous way. * Bring up to date * Add settings toggle. Fix Muiltithreading Off mode. * Fix warning. * Release tracking lock for flushes * Fix Conditional Render fast path with threaded gal * Make handle iteration safe when releasing the lock This is mostly temporary. * Attempt to set backend threading on driver Only really works on nvidia before launching a game. * Fix race condition with BufferModifiedRangeList, exceptions in tracking actions * Update buffer set commands * Some cleanup * Only use stutter workaround when using opengl renderer non-threaded * Add host-conditional reservation of counter events There has always been the possibility that conditional rendering could use a query object just as it is disposed by the counter queue. This change makes it so that when the host decides to use host conditional rendering, the query object is reserved so that it cannot be deleted. Counter events can optionally start reserved, as the threaded implementation can reserve them before the backend creates them, and there would otherwise be a short amount of time where the counter queue could dispose the event before a call to reserve it could be made. * Address Feedback * Make counter flush tracked again. Hopefully does not cause any issues this time. * Wait for FlushTo on the main queue thread. Currently assumes only one thread will want to FlushTo (in this case, the GPU thread) * Add SDL2 headless integration * Add HLE macro commands. Co-authored-by: Mary <mary@mary.zone>
2021-08-27 00:31:29 +02:00
_programsToSaveQueue = new Queue<(IProgram, Action<byte[]>)>();
}
/// <summary>
/// Processes the queue of shaders that must save their binaries to the disk cache.
/// </summary>
public void ProcessShaderCacheQueue()
{
// Check to see if the binaries for previously compiled shaders are ready, and save them out.
while (_programsToSaveQueue.Count > 0)
{
(IProgram program, Action<byte[]> dataAction) = _programsToSaveQueue.Peek();
if (program.CheckProgramLink(false) != ProgramLinkStatus.Incomplete)
{
dataAction(program.GetBinary());
_programsToSaveQueue.Dequeue();
}
else
{
break;
}
}
}
/// <summary>
/// Initialize the cache.
/// </summary>
internal void Initialize()
{
if (GraphicsConfig.EnableShaderCache && GraphicsConfig.TitleId != null)
{
_cacheManager = new CacheManager(CacheGraphicsApi.OpenGL, CacheHashType.XxHash128, "glsl", GraphicsConfig.TitleId, ShaderCodeGenVersion);
bool isReadOnly = _cacheManager.IsReadOnly;
HashSet<Hash128> invalidEntries = null;
if (isReadOnly)
{
Logger.Warning?.Print(LogClass.Gpu, "Loading shader cache in read-only mode (cache in use by another program!)");
}
else
{
invalidEntries = new HashSet<Hash128>();
}
ReadOnlySpan<Hash128> guestProgramList = _cacheManager.GetGuestProgramList();
using AutoResetEvent progressReportEvent = new AutoResetEvent(false);
_shaderCount = 0;
_totalShaderCount = guestProgramList.Length;
ShaderCacheStateChanged?.Invoke(ShaderCacheState.Start, _shaderCount, _totalShaderCount);
Thread progressReportThread = null;
if (guestProgramList.Length > 0)
{
progressReportThread = new Thread(ReportProgress)
{
Name = "ShaderCache.ProgressReporter",
Priority = ThreadPriority.Lowest,
IsBackground = true
};
progressReportThread.Start(progressReportEvent);
}
// Make sure these are initialized before doing compilation.
Capabilities caps = _context.Capabilities;
int maxTaskCount = Math.Min(Environment.ProcessorCount, 8);
int programIndex = 0;
List<ShaderCompileTask> activeTasks = new List<ShaderCompileTask>();
using AutoResetEvent taskDoneEvent = new AutoResetEvent(false);
// This thread dispatches tasks to do shader translation, and creates programs that OpenGL will link in the background.
// The program link status is checked in a non-blocking manner so that multiple shaders can be compiled at once.
while (programIndex < guestProgramList.Length || activeTasks.Count > 0)
{
if (activeTasks.Count < maxTaskCount && programIndex < guestProgramList.Length)
{
// Begin a new shader compilation.
Hash128 key = guestProgramList[programIndex];
byte[] hostProgramBinary = _cacheManager.GetHostProgramByHash(ref key);
bool hasHostCache = hostProgramBinary != null;
IProgram hostProgram = null;
// If the program sources aren't in the cache, compile from saved guest program.
byte[] guestProgram = _cacheManager.GetGuestProgramByHash(ref key);
if (guestProgram == null)
{
Logger.Error?.Print(LogClass.Gpu, $"Ignoring orphan shader hash {key} in cache (is the cache incomplete?)");
// Should not happen, but if someone messed with the cache it's better to catch it.
invalidEntries?.Add(key);
2021-04-15 00:15:25 +02:00
_shaderCount = ++programIndex;
continue;
}
ReadOnlySpan<byte> guestProgramReadOnlySpan = guestProgram;
ReadOnlySpan<GuestShaderCacheEntry> cachedShaderEntries = GuestShaderCacheEntry.Parse(ref guestProgramReadOnlySpan, out GuestShaderCacheHeader fileHeader);
if (cachedShaderEntries[0].Header.Stage == ShaderStage.Compute)
{
Debug.Assert(cachedShaderEntries.Length == 1);
GuestShaderCacheEntry entry = cachedShaderEntries[0];
HostShaderCacheEntry[] hostShaderEntries = null;
// Try loading host shader binary.
if (hasHostCache)
{
hostShaderEntries = HostShaderCacheEntry.Parse(hostProgramBinary, out ReadOnlySpan<byte> hostProgramBinarySpan);
hostProgramBinary = hostProgramBinarySpan.ToArray();
hostProgram = _context.Renderer.LoadProgramBinary(hostProgramBinary);
}
ShaderCompileTask task = new ShaderCompileTask(taskDoneEvent);
activeTasks.Add(task);
task.OnCompiled(hostProgram, (bool isHostProgramValid, ShaderCompileTask task) =>
{
ShaderProgram program = null;
ShaderProgramInfo shaderProgramInfo = null;
if (isHostProgramValid)
{
// Reconstruct code holder.
program = new ShaderProgram(entry.Header.Stage, "");
shaderProgramInfo = hostShaderEntries[0].ToShaderProgramInfo();
byte[] code = entry.Code.AsSpan().Slice(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
ShaderCodeHolder shader = new ShaderCodeHolder(program, shaderProgramInfo, code);
_cpProgramsDiskCache.Add(key, new ShaderBundle(hostProgram, shader));
return true;
}
else
{
// If the host program was rejected by the gpu driver or isn't in cache, try to build from program sources again.
Task compileTask = Task.Run(() =>
{
var binaryCode = new Memory<byte>(entry.Code);
var gpuAccessor = new CachedGpuAccessor(
_context,
binaryCode,
binaryCode.Slice(binaryCode.Length - entry.Header.Cb1DataSize),
entry.Header.GpuAccessorHeader,
entry.TextureDescriptors,
null);
var options = new TranslationOptions(TargetLanguage.Glsl, TargetApi.OpenGL, DefaultFlags | TranslationFlags.Compute);
program = Translator.CreateContext(0, gpuAccessor, options).Translate(out shaderProgramInfo);
});
task.OnTask(compileTask, (bool _, ShaderCompileTask task) =>
{
if (task.IsFaulted)
{
Logger.Warning?.Print(LogClass.Gpu, $"Host shader {key} is corrupted or incompatible, discarding...");
_cacheManager.RemoveProgram(ref key);
return true; // Exit early, the decoding step failed.
}
byte[] code = entry.Code.AsSpan().Slice(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
ShaderCodeHolder shader = new ShaderCodeHolder(program, shaderProgramInfo, code);
Logger.Info?.Print(LogClass.Gpu, $"Host shader {key} got invalidated, rebuilding from guest...");
// Compile shader and create program as the shader program binary got invalidated.
shader.HostShader = _context.Renderer.CompileShader(ShaderStage.Compute, program.Code);
hostProgram = _context.Renderer.CreateProgram(new IShader[] { shader.HostShader });
task.OnCompiled(hostProgram, (bool isNewProgramValid, ShaderCompileTask task) =>
{
// As the host program was invalidated, save the new entry in the cache.
hostProgramBinary = HostShaderCacheEntry.Create(hostProgram.GetBinary(), new ShaderCodeHolder[] { shader });
if (!isReadOnly)
{
if (hasHostCache)
{
_cacheManager.ReplaceHostProgram(ref key, hostProgramBinary);
}
else
{
Logger.Warning?.Print(LogClass.Gpu, $"Add missing host shader {key} in cache (is the cache incomplete?)");
_cacheManager.AddHostProgram(ref key, hostProgramBinary);
}
}
_cpProgramsDiskCache.Add(key, new ShaderBundle(hostProgram, shader));
return true;
});
return false; // Not finished: still need to compile the host program.
});
return false; // Not finished: translating the program.
}
});
}
else
{
Debug.Assert(cachedShaderEntries.Length == Constants.ShaderStages);
ShaderCodeHolder[] shaders = new ShaderCodeHolder[cachedShaderEntries.Length];
List<ShaderProgram> shaderPrograms = new List<ShaderProgram>();
TransformFeedbackDescriptor[] tfd = CacheHelper.ReadTransformFeedbackInformation(ref guestProgramReadOnlySpan, fileHeader);
TranslationCounts counts = new TranslationCounts();
HostShaderCacheEntry[] hostShaderEntries = null;
// Try loading host shader binary.
if (hasHostCache)
{
hostShaderEntries = HostShaderCacheEntry.Parse(hostProgramBinary, out ReadOnlySpan<byte> hostProgramBinarySpan);
hostProgramBinary = hostProgramBinarySpan.ToArray();
hostProgram = _context.Renderer.LoadProgramBinary(hostProgramBinary);
}
ShaderCompileTask task = new ShaderCompileTask(taskDoneEvent);
activeTasks.Add(task);
GuestShaderCacheEntry[] entries = cachedShaderEntries.ToArray();
task.OnCompiled(hostProgram, (bool isHostProgramValid, ShaderCompileTask task) =>
{
Task compileTask = Task.Run(() =>
{
TranslatorContext[] shaderContexts = null;
if (!isHostProgramValid)
{
shaderContexts = new TranslatorContext[1 + entries.Length];
for (int i = 0; i < entries.Length; i++)
{
GuestShaderCacheEntry entry = entries[i];
if (entry == null)
{
continue;
}
var binaryCode = new Memory<byte>(entry.Code);
var gpuAccessor = new CachedGpuAccessor(
_context,
binaryCode,
binaryCode.Slice(binaryCode.Length - entry.Header.Cb1DataSize),
entry.Header.GpuAccessorHeader,
entry.TextureDescriptors,
tfd);
var options = new TranslationOptions(TargetLanguage.Glsl, TargetApi.OpenGL, DefaultFlags);
shaderContexts[i + 1] = Translator.CreateContext(0, gpuAccessor, options, counts);
if (entry.Header.SizeA != 0)
{
var options2 = new TranslationOptions(TargetLanguage.Glsl, TargetApi.OpenGL, DefaultFlags | TranslationFlags.VertexA);
shaderContexts[0] = Translator.CreateContext((ulong)entry.Header.Size, gpuAccessor, options2, counts);
}
}
}
// Reconstruct code holder.
for (int i = 0; i < entries.Length; i++)
{
GuestShaderCacheEntry entry = entries[i];
if (entry == null)
{
continue;
}
ShaderProgram program;
ShaderProgramInfo shaderProgramInfo;
if (isHostProgramValid)
{
program = new ShaderProgram(entry.Header.Stage, "");
shaderProgramInfo = hostShaderEntries[i].ToShaderProgramInfo();
}
else
{
int stageIndex = i + 1;
TranslatorContext currentStage = shaderContexts[stageIndex];
TranslatorContext nextStage = GetNextStageContext(shaderContexts, stageIndex);
TranslatorContext vertexA = stageIndex == 1 ? shaderContexts[0] : null;
program = currentStage.Translate(out shaderProgramInfo, nextStage, vertexA);
}
// NOTE: Vertex B comes first in the shader cache.
byte[] code = entry.Code.AsSpan().Slice(0, entry.Header.Size - entry.Header.Cb1DataSize).ToArray();
byte[] code2 = entry.Header.SizeA != 0 ? entry.Code.AsSpan().Slice(entry.Header.Size, entry.Header.SizeA).ToArray() : null;
shaders[i] = new ShaderCodeHolder(program, shaderProgramInfo, code, code2);
shaderPrograms.Add(program);
}
});
task.OnTask(compileTask, (bool _, ShaderCompileTask task) =>
{
if (task.IsFaulted)
{
Logger.Warning?.Print(LogClass.Gpu, $"Host shader {key} is corrupted or incompatible, discarding...");
_cacheManager.RemoveProgram(ref key);
return true; // Exit early, the decoding step failed.
}
// If the host program was rejected by the gpu driver or isn't in cache, try to build from program sources again.
if (!isHostProgramValid)
{
Logger.Info?.Print(LogClass.Gpu, $"Host shader {key} got invalidated, rebuilding from guest...");
List<IShader> hostShaders = new List<IShader>();
// Compile shaders and create program as the shader program binary got invalidated.
for (int stage = 0; stage < Constants.ShaderStages; stage++)
{
ShaderProgram program = shaders[stage]?.Program;
if (program == null)
{
continue;
}
IShader hostShader = _context.Renderer.CompileShader(program.Stage, program.Code);
shaders[stage].HostShader = hostShader;
hostShaders.Add(hostShader);
}
hostProgram = _context.Renderer.CreateProgram(hostShaders.ToArray());
task.OnCompiled(hostProgram, (bool isNewProgramValid, ShaderCompileTask task) =>
{
// As the host program was invalidated, save the new entry in the cache.
hostProgramBinary = HostShaderCacheEntry.Create(hostProgram.GetBinary(), shaders);
if (!isReadOnly)
{
if (hasHostCache)
{
_cacheManager.ReplaceHostProgram(ref key, hostProgramBinary);
}
else
{
Logger.Warning?.Print(LogClass.Gpu, $"Add missing host shader {key} in cache (is the cache incomplete?)");
_cacheManager.AddHostProgram(ref key, hostProgramBinary);
}
}
_gpProgramsDiskCache.Add(key, new ShaderBundle(hostProgram, shaders));
return true;
});
return false; // Not finished: still need to compile the host program.
}
else
{
_gpProgramsDiskCache.Add(key, new ShaderBundle(hostProgram, shaders));
return true;
}
});
return false; // Not finished: translating the program.
});
}
_shaderCount = ++programIndex;
}
// Process the queue.
for (int i = 0; i < activeTasks.Count; i++)
{
ShaderCompileTask task = activeTasks[i];
if (task.IsDone())
{
activeTasks.RemoveAt(i--);
}
}
if (activeTasks.Count == maxTaskCount)
{
// Wait for a task to be done, or for 1ms.
// Host shader compilation cannot signal when it is done,
// so the 1ms timeout is required to poll status.
taskDoneEvent.WaitOne(1);
}
}
if (!isReadOnly)
{
// Remove entries that are broken in the cache
_cacheManager.RemoveManifestEntries(invalidEntries);
_cacheManager.FlushToArchive();
_cacheManager.Synchronize();
}
progressReportEvent.Set();
progressReportThread?.Join();
ShaderCacheStateChanged?.Invoke(ShaderCacheState.Loaded, _shaderCount, _totalShaderCount);
Logger.Info?.Print(LogClass.Gpu, $"Shader cache loaded {_shaderCount} entries.");
}
}
/// <summary>
/// Raises ShaderCacheStateChanged events periodically.
/// </summary>
private void ReportProgress(object state)
{
const int refreshRate = 50; // ms
AutoResetEvent endEvent = (AutoResetEvent)state;
int count = 0;
do
{
int newCount = _shaderCount;
if (count != newCount)
{
ShaderCacheStateChanged?.Invoke(ShaderCacheState.Loading, newCount, _totalShaderCount);
count = newCount;
}
}
while (!endEvent.WaitOne(refreshRate));
}
/// <summary>
/// Gets a compute shader from the cache.
/// </summary>
2020-01-01 16:39:09 +01:00
/// <remarks>
/// This automatically translates, compiles and adds the code to the cache if not present.
/// </remarks>
/// <param name="channel">GPU channel</param>
/// <param name="gas">GPU accessor state</param>
/// <param name="gpuVa">GPU virtual address of the binary shader code</param>
/// <param name="localSizeX">Local group size X of the computer shader</param>
/// <param name="localSizeY">Local group size Y of the computer shader</param>
/// <param name="localSizeZ">Local group size Z of the computer shader</param>
/// <param name="localMemorySize">Local memory size of the compute shader</param>
/// <param name="sharedMemorySize">Shared memory size of the compute shader</param>
/// <returns>Compiled compute shader code</returns>
public ShaderBundle GetComputeShader(
GpuChannel channel,
GpuAccessorState gas,
ulong gpuVa,
int localSizeX,
int localSizeY,
int localSizeZ,
int localMemorySize,
int sharedMemorySize)
{
bool isCached = _cpPrograms.TryGetValue(gpuVa, out List<ShaderBundle> list);
if (isCached)
{
foreach (ShaderBundle cachedCpShader in list)
{
if (IsShaderEqual(channel.MemoryManager, cachedCpShader, gpuVa))
{
return cachedCpShader;
}
}
}
TranslatorContext[] shaderContexts = new TranslatorContext[1];
shaderContexts[0] = DecodeComputeShader(
channel,
gas,
gpuVa,
localSizeX,
localSizeY,
localSizeZ,
localMemorySize,
sharedMemorySize);
bool isShaderCacheEnabled = _cacheManager != null;
bool isShaderCacheReadOnly = false;
Hash128 programCodeHash = default;
GuestShaderCacheEntry[] shaderCacheEntries = null;
// Current shader cache doesn't support bindless textures
if (shaderContexts[0].UsedFeatures.HasFlag(FeatureFlags.Bindless))
{
isShaderCacheEnabled = false;
}
if (isShaderCacheEnabled)
{
isShaderCacheReadOnly = _cacheManager.IsReadOnly;
// Compute hash and prepare data for shader disk cache comparison.
shaderCacheEntries = CacheHelper.CreateShaderCacheEntries(channel, shaderContexts);
programCodeHash = CacheHelper.ComputeGuestHashFromCache(shaderCacheEntries);
}
ShaderBundle cpShader;
// Search for the program hash in loaded shaders.
if (!isShaderCacheEnabled || !_cpProgramsDiskCache.TryGetValue(programCodeHash, out cpShader))
{
if (isShaderCacheEnabled)
{
Logger.Debug?.Print(LogClass.Gpu, $"Shader {programCodeHash} not in cache, compiling!");
}
// The shader isn't currently cached, translate it and compile it.
ShaderCodeHolder shader = TranslateShader(_dumper, channel.MemoryManager, shaderContexts[0], null, null);
shader.HostShader = _context.Renderer.CompileShader(ShaderStage.Compute, shader.Program.Code);
IProgram hostProgram = _context.Renderer.CreateProgram(new IShader[] { shader.HostShader });
cpShader = new ShaderBundle(hostProgram, shader);
if (isShaderCacheEnabled)
{
_cpProgramsDiskCache.Add(programCodeHash, cpShader);
if (!isShaderCacheReadOnly)
{
Add a Multithreading layer for the GAL, multi-thread shader compilation at runtime (#2501) * Initial Implementation About as fast as nvidia GL multithreading, can be improved with faster command queuing. * Struct based command list Speeds up a bit. Still a lot of time lost to resource copy. * Do shader init while the render thread is active. * Introduce circular span pool V1 Ideally should be able to use structs instead of references for storing these spans on commands. Will try that next. * Refactor SpanRef some more Use a struct to represent SpanRef, rather than a reference. * Flush buffers on background thread * Use a span for UpdateRenderScale. Much faster than copying the array. * Calculate command size using reflection * WIP parallel shaders * Some minor optimisation * Only 2 max refs per command now. The command with 3 refs is gone. :relieved: * Don't cast on the GPU side * Remove redundant casts, force sync on window present * Fix Shader Cache * Fix host shader save. * Fixup to work with new renderer stuff * Make command Run static, use array of delegates as lookup Profile says this takes less time than the previous way. * Bring up to date * Add settings toggle. Fix Muiltithreading Off mode. * Fix warning. * Release tracking lock for flushes * Fix Conditional Render fast path with threaded gal * Make handle iteration safe when releasing the lock This is mostly temporary. * Attempt to set backend threading on driver Only really works on nvidia before launching a game. * Fix race condition with BufferModifiedRangeList, exceptions in tracking actions * Update buffer set commands * Some cleanup * Only use stutter workaround when using opengl renderer non-threaded * Add host-conditional reservation of counter events There has always been the possibility that conditional rendering could use a query object just as it is disposed by the counter queue. This change makes it so that when the host decides to use host conditional rendering, the query object is reserved so that it cannot be deleted. Counter events can optionally start reserved, as the threaded implementation can reserve them before the backend creates them, and there would otherwise be a short amount of time where the counter queue could dispose the event before a call to reserve it could be made. * Address Feedback * Make counter flush tracked again. Hopefully does not cause any issues this time. * Wait for FlushTo on the main queue thread. Currently assumes only one thread will want to FlushTo (in this case, the GPU thread) * Add SDL2 headless integration * Add HLE macro commands. Co-authored-by: Mary <mary@mary.zone>
2021-08-27 00:31:29 +02:00
byte[] guestProgramDump = CacheHelper.CreateGuestProgramDump(shaderCacheEntries);
_programsToSaveQueue.Enqueue((hostProgram, (byte[] hostProgramBinary) =>
{
_cacheManager.SaveProgram(ref programCodeHash, guestProgramDump, HostShaderCacheEntry.Create(hostProgramBinary, new ShaderCodeHolder[] { shader }));
}));
}
}
}
if (!isCached)
{
list = new List<ShaderBundle>();
_cpPrograms.Add(gpuVa, list);
}
list.Add(cpShader);
return cpShader;
}
/// <summary>
/// Gets a graphics shader program from the shader cache.
/// This includes all the specified shader stages.
/// </summary>
2020-01-01 16:39:09 +01:00
/// <remarks>
/// This automatically translates, compiles and adds the code to the cache if not present.
/// </remarks>
/// <param name="state">GPU state</param>
/// <param name="channel">GPU channel</param>
/// <param name="gas">GPU accessor state</param>
/// <param name="addresses">Addresses of the shaders for each stage</param>
/// <returns>Compiled graphics shader code</returns>
public ShaderBundle GetGraphicsShader(ref ThreedClassState state, GpuChannel channel, GpuAccessorState gas, ShaderAddresses addresses)
{
bool isCached = _gpPrograms.TryGetValue(addresses, out List<ShaderBundle> list);
if (isCached)
{
foreach (ShaderBundle cachedGpShaders in list)
{
if (IsShaderEqual(channel.MemoryManager, cachedGpShaders, addresses))
{
return cachedGpShaders;
}
}
}
TranslatorContext[] shaderContexts = new TranslatorContext[Constants.ShaderStages + 1];
TransformFeedbackDescriptor[] tfd = GetTransformFeedbackDescriptors(ref state);
gas.TransformFeedbackDescriptors = tfd;
TranslationCounts counts = new TranslationCounts();
if (addresses.VertexA != 0)
{
shaderContexts[0] = DecodeGraphicsShader(channel, gas, counts, DefaultFlags | TranslationFlags.VertexA, ShaderStage.Vertex, addresses.VertexA);
}
shaderContexts[1] = DecodeGraphicsShader(channel, gas, counts, DefaultFlags, ShaderStage.Vertex, addresses.Vertex);
shaderContexts[2] = DecodeGraphicsShader(channel, gas, counts, DefaultFlags, ShaderStage.TessellationControl, addresses.TessControl);
shaderContexts[3] = DecodeGraphicsShader(channel, gas, counts, DefaultFlags, ShaderStage.TessellationEvaluation, addresses.TessEvaluation);
shaderContexts[4] = DecodeGraphicsShader(channel, gas, counts, DefaultFlags, ShaderStage.Geometry, addresses.Geometry);
shaderContexts[5] = DecodeGraphicsShader(channel, gas, counts, DefaultFlags, ShaderStage.Fragment, addresses.Fragment);
bool isShaderCacheEnabled = _cacheManager != null;
bool isShaderCacheReadOnly = false;
Hash128 programCodeHash = default;
GuestShaderCacheEntry[] shaderCacheEntries = null;
// Current shader cache doesn't support bindless textures
for (int i = 0; i < shaderContexts.Length; i++)
{
if (shaderContexts[i] != null && shaderContexts[i].UsedFeatures.HasFlag(FeatureFlags.Bindless))
{
isShaderCacheEnabled = false;
break;
}
}
if (isShaderCacheEnabled)
{
isShaderCacheReadOnly = _cacheManager.IsReadOnly;
// Compute hash and prepare data for shader disk cache comparison.
shaderCacheEntries = CacheHelper.CreateShaderCacheEntries(channel, shaderContexts);
programCodeHash = CacheHelper.ComputeGuestHashFromCache(shaderCacheEntries, tfd);
}
ShaderBundle gpShaders;
// Search for the program hash in loaded shaders.
if (!isShaderCacheEnabled || !_gpProgramsDiskCache.TryGetValue(programCodeHash, out gpShaders))
{
if (isShaderCacheEnabled)
{
Logger.Debug?.Print(LogClass.Gpu, $"Shader {programCodeHash} not in cache, compiling!");
}
// The shader isn't currently cached, translate it and compile it.
ShaderCodeHolder[] shaders = new ShaderCodeHolder[Constants.ShaderStages];
for (int stageIndex = 0; stageIndex < Constants.ShaderStages; stageIndex++)
{
shaders[stageIndex] = TranslateShader(_dumper, channel.MemoryManager, shaderContexts, stageIndex + 1);
}
List<IShader> hostShaders = new List<IShader>();
for (int stage = 0; stage < Constants.ShaderStages; stage++)
{
ShaderProgram program = shaders[stage]?.Program;
if (program == null)
{
continue;
}
IShader hostShader = _context.Renderer.CompileShader(program.Stage, program.Code);
shaders[stage].HostShader = hostShader;
hostShaders.Add(hostShader);
}
IProgram hostProgram = _context.Renderer.CreateProgram(hostShaders.ToArray());
gpShaders = new ShaderBundle(hostProgram, shaders);
if (isShaderCacheEnabled)
{
_gpProgramsDiskCache.Add(programCodeHash, gpShaders);
if (!isShaderCacheReadOnly)
{
Add a Multithreading layer for the GAL, multi-thread shader compilation at runtime (#2501) * Initial Implementation About as fast as nvidia GL multithreading, can be improved with faster command queuing. * Struct based command list Speeds up a bit. Still a lot of time lost to resource copy. * Do shader init while the render thread is active. * Introduce circular span pool V1 Ideally should be able to use structs instead of references for storing these spans on commands. Will try that next. * Refactor SpanRef some more Use a struct to represent SpanRef, rather than a reference. * Flush buffers on background thread * Use a span for UpdateRenderScale. Much faster than copying the array. * Calculate command size using reflection * WIP parallel shaders * Some minor optimisation * Only 2 max refs per command now. The command with 3 refs is gone. :relieved: * Don't cast on the GPU side * Remove redundant casts, force sync on window present * Fix Shader Cache * Fix host shader save. * Fixup to work with new renderer stuff * Make command Run static, use array of delegates as lookup Profile says this takes less time than the previous way. * Bring up to date * Add settings toggle. Fix Muiltithreading Off mode. * Fix warning. * Release tracking lock for flushes * Fix Conditional Render fast path with threaded gal * Make handle iteration safe when releasing the lock This is mostly temporary. * Attempt to set backend threading on driver Only really works on nvidia before launching a game. * Fix race condition with BufferModifiedRangeList, exceptions in tracking actions * Update buffer set commands * Some cleanup * Only use stutter workaround when using opengl renderer non-threaded * Add host-conditional reservation of counter events There has always been the possibility that conditional rendering could use a query object just as it is disposed by the counter queue. This change makes it so that when the host decides to use host conditional rendering, the query object is reserved so that it cannot be deleted. Counter events can optionally start reserved, as the threaded implementation can reserve them before the backend creates them, and there would otherwise be a short amount of time where the counter queue could dispose the event before a call to reserve it could be made. * Address Feedback * Make counter flush tracked again. Hopefully does not cause any issues this time. * Wait for FlushTo on the main queue thread. Currently assumes only one thread will want to FlushTo (in this case, the GPU thread) * Add SDL2 headless integration * Add HLE macro commands. Co-authored-by: Mary <mary@mary.zone>
2021-08-27 00:31:29 +02:00
byte[] guestProgramDump = CacheHelper.CreateGuestProgramDump(shaderCacheEntries, tfd);
_programsToSaveQueue.Enqueue((hostProgram, (byte[] hostProgramBinary) =>
{
_cacheManager.SaveProgram(ref programCodeHash, guestProgramDump, HostShaderCacheEntry.Create(hostProgramBinary, shaders));
}));
}
}
}
if (!isCached)
{
list = new List<ShaderBundle>();
_gpPrograms.Add(addresses, list);
}
list.Add(gpShaders);
return gpShaders;
}
/// <summary>
/// Gets transform feedback state from the current GPU state.
/// </summary>
/// <param name="state">Current GPU state</param>
/// <returns>Four transform feedback descriptors for the enabled TFBs, or null if TFB is disabled</returns>
private static TransformFeedbackDescriptor[] GetTransformFeedbackDescriptors(ref ThreedClassState state)
{
bool tfEnable = state.TfEnable;
if (!tfEnable)
{
return null;
}
TransformFeedbackDescriptor[] descs = new TransformFeedbackDescriptor[Constants.TotalTransformFeedbackBuffers];
for (int i = 0; i < Constants.TotalTransformFeedbackBuffers; i++)
{
var tf = state.TfState[i];
int length = (int)Math.Min((uint)tf.VaryingsCount, 0x80);
var varyingLocations = MemoryMarshal.Cast<uint, byte>(state.TfVaryingLocations[i].ToSpan()).Slice(0, length);
descs[i] = new TransformFeedbackDescriptor(tf.BufferIndex, tf.Stride, varyingLocations.ToArray());
}
return descs;
}
/// <summary>
/// Checks if compute shader code in memory is equal to the cached shader.
/// </summary>
/// <param name="memoryManager">Memory manager used to access the GPU memory where the shader is located</param>
/// <param name="cpShader">Cached compute shader</param>
/// <param name="gpuVa">GPU virtual address of the shader code in memory</param>
/// <returns>True if the code is different, false otherwise</returns>
private static bool IsShaderEqual(MemoryManager memoryManager, ShaderBundle cpShader, ulong gpuVa)
{
return IsShaderEqual(memoryManager, cpShader.Shaders[0], gpuVa);
}
/// <summary>
/// Checks if graphics shader code from all stages in memory are equal to the cached shaders.
/// </summary>
/// <param name="memoryManager">Memory manager used to access the GPU memory where the shader is located</param>
/// <param name="gpShaders">Cached graphics shaders</param>
/// <param name="addresses">GPU virtual addresses of all enabled shader stages</param>
/// <returns>True if the code is different, false otherwise</returns>
private static bool IsShaderEqual(MemoryManager memoryManager, ShaderBundle gpShaders, ShaderAddresses addresses)
{
2019-12-31 23:09:49 +01:00
for (int stage = 0; stage < gpShaders.Shaders.Length; stage++)
{
ShaderCodeHolder shader = gpShaders.Shaders[stage];
ulong gpuVa = 0;
switch (stage)
{
case 0: gpuVa = addresses.Vertex; break;
case 1: gpuVa = addresses.TessControl; break;
case 2: gpuVa = addresses.TessEvaluation; break;
case 3: gpuVa = addresses.Geometry; break;
case 4: gpuVa = addresses.Fragment; break;
}
if (!IsShaderEqual(memoryManager, shader, gpuVa, addresses.VertexA))
{
return false;
}
}
return true;
}
/// <summary>
/// Checks if the code of the specified cached shader is different from the code in memory.
/// </summary>
/// <param name="memoryManager">Memory manager used to access the GPU memory where the shader is located</param>
/// <param name="shader">Cached shader to compare with</param>
/// <param name="gpuVa">GPU virtual address of the binary shader code</param>
/// <param name="gpuVaA">Optional GPU virtual address of the "Vertex A" binary shader code</param>
/// <returns>True if the code is different, false otherwise</returns>
private static bool IsShaderEqual(MemoryManager memoryManager, ShaderCodeHolder shader, ulong gpuVa, ulong gpuVaA = 0)
{
2020-01-05 17:40:21 +01:00
if (shader == null)
{
return true;
2020-01-05 17:40:21 +01:00
}
ReadOnlySpan<byte> memoryCode = memoryManager.GetSpan(gpuVa, shader.Code.Length);
bool equals = memoryCode.SequenceEqual(shader.Code);
if (equals && shader.Code2 != null)
{
memoryCode = memoryManager.GetSpan(gpuVaA, shader.Code2.Length);
equals = memoryCode.SequenceEqual(shader.Code2);
}
return equals;
}
/// <summary>
/// Decode the binary Maxwell shader code to a translator context.
/// </summary>
/// <param name="channel">GPU channel</param>
/// <param name="gas">GPU accessor state</param>
/// <param name="gpuVa">GPU virtual address of the binary shader code</param>
/// <param name="localSizeX">Local group size X of the computer shader</param>
/// <param name="localSizeY">Local group size Y of the computer shader</param>
/// <param name="localSizeZ">Local group size Z of the computer shader</param>
/// <param name="localMemorySize">Local memory size of the compute shader</param>
/// <param name="sharedMemorySize">Shared memory size of the compute shader</param>
/// <returns>The generated translator context</returns>
private TranslatorContext DecodeComputeShader(
GpuChannel channel,
GpuAccessorState gas,
ulong gpuVa,
int localSizeX,
int localSizeY,
int localSizeZ,
int localMemorySize,
int sharedMemorySize)
{
if (gpuVa == 0)
{
return null;
}
GpuAccessor gpuAccessor = new GpuAccessor(_context, channel, gas, localSizeX, localSizeY, localSizeZ, localMemorySize, sharedMemorySize);
var options = new TranslationOptions(TargetLanguage.Glsl, TargetApi.OpenGL, DefaultFlags | TranslationFlags.Compute);
return Translator.CreateContext(gpuVa, gpuAccessor, options);
}
/// <summary>
/// Decode the binary Maxwell shader code to a translator context.
/// </summary>
2020-01-01 16:39:09 +01:00
/// <remarks>
/// This will combine the "Vertex A" and "Vertex B" shader stages, if specified, into one shader.
/// </remarks>
/// <param name="channel">GPU channel</param>
/// <param name="gas">GPU accessor state</param>
/// <param name="counts">Cumulative shader resource counts</param>
/// <param name="flags">Flags that controls shader translation</param>
/// <param name="stage">Shader stage</param>
/// <param name="gpuVa">GPU virtual address of the shader code</param>
/// <returns>The generated translator context</returns>
private TranslatorContext DecodeGraphicsShader(
GpuChannel channel,
GpuAccessorState gas,
TranslationCounts counts,
TranslationFlags flags,
ShaderStage stage,
ulong gpuVa)
{
if (gpuVa == 0)
{
2020-01-05 17:40:21 +01:00
return null;
}
GpuAccessor gpuAccessor = new GpuAccessor(_context, channel, gas, (int)stage - 1);
var options = new TranslationOptions(TargetLanguage.Glsl, TargetApi.OpenGL, flags);
return Translator.CreateContext(gpuVa, gpuAccessor, options, counts);
}
/// <summary>
/// Translates a previously generated translator context to something that the host API accepts.
/// </summary>
/// <param name="dumper">Optional shader code dumper</param>
/// <param name="memoryManager">Memory manager used to access the GPU memory where the shader is located</param>
/// <param name="stages">Translator context of all available shader stages</param>
/// <param name="stageIndex">Index on the stages array to translate</param>
/// <returns>Compiled graphics shader code</returns>
private static ShaderCodeHolder TranslateShader(
ShaderDumper dumper,
MemoryManager memoryManager,
TranslatorContext[] stages,
int stageIndex)
{
TranslatorContext currentStage = stages[stageIndex];
TranslatorContext nextStage = GetNextStageContext(stages, stageIndex);
TranslatorContext vertexA = stageIndex == 1 ? stages[0] : null;
return TranslateShader(dumper, memoryManager, currentStage, nextStage, vertexA);
}
/// <summary>
/// Gets the next shader stage context, from an array of contexts and index of the current stage.
/// </summary>
/// <param name="stages">Translator context of all available shader stages</param>
/// <param name="stageIndex">Index on the stages array to translate</param>
/// <returns>The translator context of the next stage, or null if inexistent</returns>
private static TranslatorContext GetNextStageContext(TranslatorContext[] stages, int stageIndex)
{
for (int nextStageIndex = stageIndex + 1; nextStageIndex < stages.Length; nextStageIndex++)
{
if (stages[nextStageIndex] != null)
{
return stages[nextStageIndex];
}
}
return null;
}
/// <summary>
/// Translates a previously generated translator context to something that the host API accepts.
/// </summary>
/// <param name="dumper">Optional shader code dumper</param>
/// <param name="memoryManager">Memory manager used to access the GPU memory where the shader is located</param>
/// <param name="currentStage">Translator context of the stage to be translated</param>
/// <param name="nextStage">Translator context of the next active stage, if existent</param>
/// <param name="vertexA">Optional translator context of the shader that should be combined</param>
/// <returns>Compiled graphics shader code</returns>
private static ShaderCodeHolder TranslateShader(
ShaderDumper dumper,
MemoryManager memoryManager,
TranslatorContext currentStage,
TranslatorContext nextStage,
TranslatorContext vertexA)
{
if (currentStage == null)
{
return null;
}
if (vertexA != null)
{
byte[] codeA = memoryManager.GetSpan(vertexA.Address, vertexA.Size).ToArray();
byte[] codeB = memoryManager.GetSpan(currentStage.Address, currentStage.Size).ToArray();
ShaderDumpPaths pathsA = default;
ShaderDumpPaths pathsB = default;
if (dumper != null)
{
pathsA = dumper.Dump(codeA, compute: false);
pathsB = dumper.Dump(codeB, compute: false);
}
ShaderProgram program = currentStage.Translate(out ShaderProgramInfo shaderProgramInfo, nextStage, vertexA);
pathsB.Prepend(program);
pathsA.Prepend(program);
return new ShaderCodeHolder(program, shaderProgramInfo, codeB, codeA);
}
else
{
byte[] code = memoryManager.GetSpan(currentStage.Address, currentStage.Size).ToArray();
ShaderDumpPaths paths = dumper?.Dump(code, currentStage.Stage == ShaderStage.Compute) ?? default;
ShaderProgram program = currentStage.Translate(out ShaderProgramInfo shaderProgramInfo, nextStage);
paths.Prepend(program);
return new ShaderCodeHolder(program, shaderProgramInfo, code);
}
}
2019-12-31 23:09:49 +01:00
/// <summary>
/// Disposes the shader cache, deleting all the cached shaders.
/// It's an error to use the shader cache after disposal.
/// </summary>
public void Dispose()
{
foreach (List<ShaderBundle> list in _cpPrograms.Values)
2019-12-31 23:09:49 +01:00
{
foreach (ShaderBundle bundle in list)
2019-12-31 23:09:49 +01:00
{
bundle.Dispose();
2019-12-31 23:09:49 +01:00
}
}
foreach (List<ShaderBundle> list in _gpPrograms.Values)
2019-12-31 23:09:49 +01:00
{
foreach (ShaderBundle bundle in list)
2019-12-31 23:09:49 +01:00
{
bundle.Dispose();
2019-12-31 23:09:49 +01:00
}
}
_cacheManager?.Dispose();
}
}
}