rjx-mirror/Ryujinx.Graphics.Vulkan/BufferHolder.cs
riperiperi 9f1cf6458c
Vulkan: Migrate buffers between memory types to improve GPU performance (#4540)
* Initial implementation of migration between memory heaps

- Missing OOM handling
- Missing `_map` data safety when remapping
  - Copy may not have completed yet (needs some kind of fence)
  - Map may be unmapped before it is done being used. (needs scoped access)
- SSBO accesses are all "writes" - maybe pass info in another way.
- Missing keeping map type when resizing buffers (should this be done?)

* Ensure migrated data is in place before flushing.

* Fix issue where old waitable would be signalled.

- There is a real issue where existing Auto<> references need to be replaced.

* Swap bound Auto<> instances when swapping buffer backing

* Fix conversion buffers

* Don't try move buffers if the host has shared memory.

* Make GPU methods return PinnedSpan with scope

* Storage Hint

* Fix stupidity

* Fix rebase

* Tweak rules

Attempt to sidestep BOTW slowdown

* Remove line

* Migrate only when command buffers flush

* Change backing swap log to debug

* Address some feedback

* Disallow backing swap when the flush lock is held by the current thread

* Make PinnedSpan from ReadOnlySpan explicitly unsafe

* Fix some small issues

- Index buffer swap fixed
- Allocate DeviceLocal buffers using a separate block list to images.

* Remove alternative flags

* Address feedback
2023-03-19 17:56:48 -03:00

788 lines
27 KiB
C#

using Ryujinx.Common.Logging;
using Ryujinx.Graphics.GAL;
using Silk.NET.Vulkan;
using System;
using System.Collections.Generic;
using System.Runtime.CompilerServices;
using System.Threading;
using VkBuffer = Silk.NET.Vulkan.Buffer;
using VkFormat = Silk.NET.Vulkan.Format;
namespace Ryujinx.Graphics.Vulkan
{
class BufferHolder : IDisposable
{
private const int MaxUpdateBufferSize = 0x10000;
private const int SetCountThreshold = 100;
private const int WriteCountThreshold = 50;
private const int FlushCountThreshold = 5;
public const int DeviceLocalSizeThreshold = 256 * 1024; // 256kb
public const AccessFlags DefaultAccessFlags =
AccessFlags.IndirectCommandReadBit |
AccessFlags.ShaderReadBit |
AccessFlags.ShaderWriteBit |
AccessFlags.TransferReadBit |
AccessFlags.TransferWriteBit |
AccessFlags.UniformReadBit;
private readonly VulkanRenderer _gd;
private readonly Device _device;
private MemoryAllocation _allocation;
private Auto<DisposableBuffer> _buffer;
private Auto<MemoryAllocation> _allocationAuto;
private ulong _bufferHandle;
private CacheByRange<BufferHolder> _cachedConvertedBuffers;
public int Size { get; }
private IntPtr _map;
private MultiFenceHolder _waitable;
private bool _lastAccessIsWrite;
private BufferAllocationType _baseType;
private BufferAllocationType _currentType;
private bool _swapQueued;
public BufferAllocationType DesiredType { get; private set; }
private int _setCount;
private int _writeCount;
private int _flushCount;
private int _flushTemp;
private ReaderWriterLock _flushLock;
private FenceHolder _flushFence;
private int _flushWaiting;
private List<Action> _swapActions;
public BufferHolder(VulkanRenderer gd, Device device, VkBuffer buffer, MemoryAllocation allocation, int size, BufferAllocationType type, BufferAllocationType currentType)
{
_gd = gd;
_device = device;
_allocation = allocation;
_allocationAuto = new Auto<MemoryAllocation>(allocation);
_waitable = new MultiFenceHolder(size);
_buffer = new Auto<DisposableBuffer>(new DisposableBuffer(gd.Api, device, buffer), _waitable, _allocationAuto);
_bufferHandle = buffer.Handle;
Size = size;
_map = allocation.HostPointer;
_baseType = type;
_currentType = currentType;
DesiredType = currentType;
_flushLock = new ReaderWriterLock();
}
public bool TryBackingSwap(ref CommandBufferScoped? cbs)
{
if (_swapQueued && DesiredType != _currentType)
{
// Only swap if the buffer is not used in any queued command buffer.
bool isRented = _buffer.HasRentedCommandBufferDependency(_gd.CommandBufferPool);
if (!isRented && _gd.CommandBufferPool.OwnedByCurrentThread && !_flushLock.IsReaderLockHeld)
{
var currentAllocation = _allocationAuto;
var currentBuffer = _buffer;
IntPtr currentMap = _map;
(VkBuffer buffer, MemoryAllocation allocation, BufferAllocationType resultType) = _gd.BufferManager.CreateBacking(_gd, Size, DesiredType, false, _currentType);
if (buffer.Handle != 0)
{
_flushLock.AcquireWriterLock(Timeout.Infinite);
ClearFlushFence();
_waitable = new MultiFenceHolder(Size);
_allocation = allocation;
_allocationAuto = new Auto<MemoryAllocation>(allocation);
_buffer = new Auto<DisposableBuffer>(new DisposableBuffer(_gd.Api, _device, buffer), _waitable, _allocationAuto);
_bufferHandle = buffer.Handle;
_map = allocation.HostPointer;
if (_map != IntPtr.Zero && currentMap != IntPtr.Zero)
{
// Copy data directly. Readbacks don't have to wait if this is done.
unsafe
{
new Span<byte>((void*)currentMap, Size).CopyTo(new Span<byte>((void*)_map, Size));
}
}
else
{
if (cbs == null)
{
cbs = _gd.CommandBufferPool.Rent();
}
CommandBufferScoped cbsV = cbs.Value;
Copy(_gd, cbsV, currentBuffer, _buffer, 0, 0, Size);
// Need to wait for the data to reach the new buffer before data can be flushed.
_flushFence = _gd.CommandBufferPool.GetFence(cbsV.CommandBufferIndex);
_flushFence.Get();
}
Logger.Debug?.PrintMsg(LogClass.Gpu, $"Converted {Size} buffer {_currentType} to {resultType}");
_currentType = resultType;
if (_swapActions != null)
{
foreach (var action in _swapActions)
{
action();
}
_swapActions.Clear();
}
currentBuffer.Dispose();
currentAllocation.Dispose();
_gd.PipelineInternal.SwapBuffer(currentBuffer, _buffer);
_flushLock.ReleaseWriterLock();
}
_swapQueued = false;
return true;
}
else
{
return false;
}
}
else
{
_swapQueued = false;
return true;
}
}
private void ConsiderBackingSwap()
{
if (_baseType == BufferAllocationType.Auto)
{
if (_writeCount >= WriteCountThreshold || _setCount >= SetCountThreshold || _flushCount >= FlushCountThreshold)
{
if (_flushCount > 0 || _flushTemp-- > 0)
{
// Buffers that flush should ideally be mapped in host address space for easy copies.
// If the buffer is large it will do better on GPU memory, as there will be more writes than data flushes (typically individual pages).
// If it is small, then it's likely most of the buffer will be flushed so we want it on host memory, as access is cached.
DesiredType = Size > DeviceLocalSizeThreshold ? BufferAllocationType.DeviceLocalMapped : BufferAllocationType.HostMapped;
// It's harder for a buffer that is flushed to revert to another type of mapping.
if (_flushCount > 0)
{
_flushTemp = 1000;
}
}
else if (_writeCount >= WriteCountThreshold)
{
// Buffers that are written often should ideally be in the device local heap. (Storage buffers)
DesiredType = BufferAllocationType.DeviceLocal;
}
else if (_setCount > SetCountThreshold)
{
// Buffers that have their data set often should ideally be host mapped. (Constant buffers)
DesiredType = BufferAllocationType.HostMapped;
}
_flushCount = 0;
_writeCount = 0;
_setCount = 0;
}
if (!_swapQueued && DesiredType != _currentType)
{
_swapQueued = true;
_gd.PipelineInternal.AddBackingSwap(this);
}
}
}
public unsafe Auto<DisposableBufferView> CreateView(VkFormat format, int offset, int size, Action invalidateView)
{
var bufferViewCreateInfo = new BufferViewCreateInfo()
{
SType = StructureType.BufferViewCreateInfo,
Buffer = new VkBuffer(_bufferHandle),
Format = format,
Offset = (uint)offset,
Range = (uint)size
};
_gd.Api.CreateBufferView(_device, bufferViewCreateInfo, null, out var bufferView).ThrowOnError();
(_swapActions ??= new List<Action>()).Add(invalidateView);
return new Auto<DisposableBufferView>(new DisposableBufferView(_gd.Api, _device, bufferView), _waitable, _buffer);
}
public void InheritMetrics(BufferHolder other)
{
_setCount = other._setCount;
_writeCount = other._writeCount;
_flushCount = other._flushCount;
_flushTemp = other._flushTemp;
}
public unsafe void InsertBarrier(CommandBuffer commandBuffer, bool isWrite)
{
// If the last access is write, we always need a barrier to be sure we will read or modify
// the correct data.
// If the last access is read, and current one is a write, we need to wait until the
// read finishes to avoid overwriting data still in use.
// Otherwise, if the last access is a read and the current one too, we don't need barriers.
bool needsBarrier = isWrite || _lastAccessIsWrite;
_lastAccessIsWrite = isWrite;
if (needsBarrier)
{
MemoryBarrier memoryBarrier = new MemoryBarrier()
{
SType = StructureType.MemoryBarrier,
SrcAccessMask = DefaultAccessFlags,
DstAccessMask = DefaultAccessFlags
};
_gd.Api.CmdPipelineBarrier(
commandBuffer,
PipelineStageFlags.AllCommandsBit,
PipelineStageFlags.AllCommandsBit,
DependencyFlags.DeviceGroupBit,
1,
memoryBarrier,
0,
null,
0,
null);
}
}
public Auto<DisposableBuffer> GetBuffer()
{
return _buffer;
}
public Auto<DisposableBuffer> GetBuffer(CommandBuffer commandBuffer, bool isWrite = false, bool isSSBO = false)
{
if (isWrite)
{
_writeCount++;
SignalWrite(0, Size);
}
else if (isSSBO)
{
// Always consider SSBO access for swapping to device local memory.
_writeCount++;
ConsiderBackingSwap();
}
return _buffer;
}
public Auto<DisposableBuffer> GetBuffer(CommandBuffer commandBuffer, int offset, int size, bool isWrite = false)
{
if (isWrite)
{
_writeCount++;
SignalWrite(offset, size);
}
return _buffer;
}
public void SignalWrite(int offset, int size)
{
ConsiderBackingSwap();
if (offset == 0 && size == Size)
{
_cachedConvertedBuffers.Clear();
}
else
{
_cachedConvertedBuffers.ClearRange(offset, size);
}
}
public BufferHandle GetHandle()
{
var handle = _bufferHandle;
return Unsafe.As<ulong, BufferHandle>(ref handle);
}
public unsafe IntPtr Map(int offset, int mappingSize)
{
return _map;
}
private void ClearFlushFence()
{
// Asusmes _flushLock is held as writer.
if (_flushFence != null)
{
if (_flushWaiting == 0)
{
_flushFence.Put();
}
_flushFence = null;
}
}
private void WaitForFlushFence()
{
// Assumes the _flushLock is held as reader, returns in same state.
if (_flushFence != null)
{
// If storage has changed, make sure the fence has been reached so that the data is in place.
var cookie = _flushLock.UpgradeToWriterLock(Timeout.Infinite);
if (_flushFence != null)
{
var fence = _flushFence;
Interlocked.Increment(ref _flushWaiting);
// Don't wait in the lock.
var restoreCookie = _flushLock.ReleaseLock();
fence.Wait();
_flushLock.RestoreLock(ref restoreCookie);
if (Interlocked.Decrement(ref _flushWaiting) == 0)
{
fence.Put();
}
_flushFence = null;
}
_flushLock.DowngradeFromWriterLock(ref cookie);
}
}
public unsafe PinnedSpan<byte> GetData(int offset, int size)
{
_flushLock.AcquireReaderLock(Timeout.Infinite);
WaitForFlushFence();
_flushCount++;
Span<byte> result;
if (_map != IntPtr.Zero)
{
result = GetDataStorage(offset, size);
// Need to be careful here, the buffer can't be unmapped while the data is being used.
_buffer.IncrementReferenceCount();
_flushLock.ReleaseReaderLock();
return PinnedSpan<byte>.UnsafeFromSpan(result, _buffer.DecrementReferenceCount);
}
else
{
BackgroundResource resource = _gd.BackgroundResources.Get();
if (_gd.CommandBufferPool.OwnedByCurrentThread)
{
_gd.FlushAllCommands();
result = resource.GetFlushBuffer().GetBufferData(_gd.CommandBufferPool, this, offset, size);
}
else
{
result = resource.GetFlushBuffer().GetBufferData(resource.GetPool(), this, offset, size);
}
_flushLock.ReleaseReaderLock();
// Flush buffer is pinned until the next GetBufferData on the thread, which is fine for current uses.
return PinnedSpan<byte>.UnsafeFromSpan(result);
}
}
public unsafe Span<byte> GetDataStorage(int offset, int size)
{
int mappingSize = Math.Min(size, Size - offset);
if (_map != IntPtr.Zero)
{
return new Span<byte>((void*)(_map + offset), mappingSize);
}
throw new InvalidOperationException("The buffer is not host mapped.");
}
public unsafe void SetData(int offset, ReadOnlySpan<byte> data, CommandBufferScoped? cbs = null, Action endRenderPass = null)
{
int dataSize = Math.Min(data.Length, Size - offset);
if (dataSize == 0)
{
return;
}
_setCount++;
if (_map != IntPtr.Zero)
{
// If persistently mapped, set the data directly if the buffer is not currently in use.
bool isRented = _buffer.HasRentedCommandBufferDependency(_gd.CommandBufferPool);
// If the buffer is rented, take a little more time and check if the use overlaps this handle.
bool needsFlush = isRented && _waitable.IsBufferRangeInUse(offset, dataSize);
if (!needsFlush)
{
WaitForFences(offset, dataSize);
data.Slice(0, dataSize).CopyTo(new Span<byte>((void*)(_map + offset), dataSize));
SignalWrite(offset, dataSize);
return;
}
}
if (cbs != null &&
_gd.PipelineInternal.RenderPassActive &&
!(_buffer.HasCommandBufferDependency(cbs.Value) &&
_waitable.IsBufferRangeInUse(cbs.Value.CommandBufferIndex, offset, dataSize)))
{
// If the buffer hasn't been used on the command buffer yet, try to preload the data.
// This avoids ending and beginning render passes on each buffer data upload.
cbs = _gd.PipelineInternal.GetPreloadCommandBuffer();
endRenderPass = null;
}
if (cbs == null ||
!VulkanConfiguration.UseFastBufferUpdates ||
data.Length > MaxUpdateBufferSize ||
!TryPushData(cbs.Value, endRenderPass, offset, data))
{
_gd.BufferManager.StagingBuffer.PushData(_gd.CommandBufferPool, cbs, endRenderPass, this, offset, data);
}
}
public unsafe void SetDataUnchecked(int offset, ReadOnlySpan<byte> data)
{
int dataSize = Math.Min(data.Length, Size - offset);
if (dataSize == 0)
{
return;
}
if (_map != IntPtr.Zero)
{
data.Slice(0, dataSize).CopyTo(new Span<byte>((void*)(_map + offset), dataSize));
}
else
{
_gd.BufferManager.StagingBuffer.PushData(_gd.CommandBufferPool, null, null, this, offset, data);
}
}
public void SetDataInline(CommandBufferScoped cbs, Action endRenderPass, int dstOffset, ReadOnlySpan<byte> data)
{
if (!TryPushData(cbs, endRenderPass, dstOffset, data))
{
throw new ArgumentException($"Invalid offset 0x{dstOffset:X} or data size 0x{data.Length:X}.");
}
}
private unsafe bool TryPushData(CommandBufferScoped cbs, Action endRenderPass, int dstOffset, ReadOnlySpan<byte> data)
{
if ((dstOffset & 3) != 0 || (data.Length & 3) != 0)
{
return false;
}
endRenderPass?.Invoke();
var dstBuffer = GetBuffer(cbs.CommandBuffer, dstOffset, data.Length, true).Get(cbs, dstOffset, data.Length).Value;
_writeCount--;
InsertBufferBarrier(
_gd,
cbs.CommandBuffer,
dstBuffer,
DefaultAccessFlags,
AccessFlags.TransferWriteBit,
PipelineStageFlags.AllCommandsBit,
PipelineStageFlags.TransferBit,
dstOffset,
data.Length);
fixed (byte* pData = data)
{
for (ulong offset = 0; offset < (ulong)data.Length;)
{
ulong size = Math.Min(MaxUpdateBufferSize, (ulong)data.Length - offset);
_gd.Api.CmdUpdateBuffer(cbs.CommandBuffer, dstBuffer, (ulong)dstOffset + offset, size, pData + offset);
offset += size;
}
}
InsertBufferBarrier(
_gd,
cbs.CommandBuffer,
dstBuffer,
AccessFlags.TransferWriteBit,
DefaultAccessFlags,
PipelineStageFlags.TransferBit,
PipelineStageFlags.AllCommandsBit,
dstOffset,
data.Length);
return true;
}
public static unsafe void Copy(
VulkanRenderer gd,
CommandBufferScoped cbs,
Auto<DisposableBuffer> src,
Auto<DisposableBuffer> dst,
int srcOffset,
int dstOffset,
int size)
{
var srcBuffer = src.Get(cbs, srcOffset, size).Value;
var dstBuffer = dst.Get(cbs, dstOffset, size).Value;
InsertBufferBarrier(
gd,
cbs.CommandBuffer,
dstBuffer,
DefaultAccessFlags,
AccessFlags.TransferWriteBit,
PipelineStageFlags.AllCommandsBit,
PipelineStageFlags.TransferBit,
dstOffset,
size);
var region = new BufferCopy((ulong)srcOffset, (ulong)dstOffset, (ulong)size);
gd.Api.CmdCopyBuffer(cbs.CommandBuffer, srcBuffer, dstBuffer, 1, &region);
InsertBufferBarrier(
gd,
cbs.CommandBuffer,
dstBuffer,
AccessFlags.TransferWriteBit,
DefaultAccessFlags,
PipelineStageFlags.TransferBit,
PipelineStageFlags.AllCommandsBit,
dstOffset,
size);
}
public static unsafe void InsertBufferBarrier(
VulkanRenderer gd,
CommandBuffer commandBuffer,
VkBuffer buffer,
AccessFlags srcAccessMask,
AccessFlags dstAccessMask,
PipelineStageFlags srcStageMask,
PipelineStageFlags dstStageMask,
int offset,
int size)
{
BufferMemoryBarrier memoryBarrier = new BufferMemoryBarrier()
{
SType = StructureType.BufferMemoryBarrier,
SrcAccessMask = srcAccessMask,
DstAccessMask = dstAccessMask,
SrcQueueFamilyIndex = Vk.QueueFamilyIgnored,
DstQueueFamilyIndex = Vk.QueueFamilyIgnored,
Buffer = buffer,
Offset = (ulong)offset,
Size = (ulong)size
};
gd.Api.CmdPipelineBarrier(
commandBuffer,
srcStageMask,
dstStageMask,
0,
0,
null,
1,
memoryBarrier,
0,
null);
}
public void WaitForFences()
{
_waitable.WaitForFences(_gd.Api, _device);
}
public void WaitForFences(int offset, int size)
{
_waitable.WaitForFences(_gd.Api, _device, offset, size);
}
private bool BoundToRange(int offset, ref int size)
{
if (offset >= Size)
{
return false;
}
size = Math.Min(Size - offset, size);
return true;
}
public Auto<DisposableBuffer> GetBufferI8ToI16(CommandBufferScoped cbs, int offset, int size)
{
if (!BoundToRange(offset, ref size))
{
return null;
}
var key = new I8ToI16CacheKey(_gd);
if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
{
holder = _gd.BufferManager.Create(_gd, (size * 2 + 3) & ~3);
_gd.PipelineInternal.EndRenderPass();
_gd.HelperShader.ConvertI8ToI16(_gd, cbs, this, holder, offset, size);
key.SetBuffer(holder.GetBuffer());
_cachedConvertedBuffers.Add(offset, size, key, holder);
}
return holder.GetBuffer();
}
public Auto<DisposableBuffer> GetAlignedVertexBuffer(CommandBufferScoped cbs, int offset, int size, int stride, int alignment)
{
if (!BoundToRange(offset, ref size))
{
return null;
}
var key = new AlignedVertexBufferCacheKey(_gd, stride, alignment);
if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
{
int alignedStride = (stride + (alignment - 1)) & -alignment;
holder = _gd.BufferManager.Create(_gd, (size / stride) * alignedStride);
_gd.PipelineInternal.EndRenderPass();
_gd.HelperShader.ChangeStride(_gd, cbs, this, holder, offset, size, stride, alignedStride);
key.SetBuffer(holder.GetBuffer());
_cachedConvertedBuffers.Add(offset, size, key, holder);
}
return holder.GetBuffer();
}
public Auto<DisposableBuffer> GetBufferTopologyConversion(CommandBufferScoped cbs, int offset, int size, IndexBufferPattern pattern, int indexSize)
{
if (!BoundToRange(offset, ref size))
{
return null;
}
var key = new TopologyConversionCacheKey(_gd, pattern, indexSize);
if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
{
// The destination index size is always I32.
int indexCount = size / indexSize;
int convertedCount = pattern.GetConvertedCount(indexCount);
holder = _gd.BufferManager.Create(_gd, convertedCount * 4);
_gd.PipelineInternal.EndRenderPass();
_gd.HelperShader.ConvertIndexBuffer(_gd, cbs, this, holder, pattern, indexSize, offset, indexCount);
key.SetBuffer(holder.GetBuffer());
_cachedConvertedBuffers.Add(offset, size, key, holder);
}
return holder.GetBuffer();
}
public bool TryGetCachedConvertedBuffer(int offset, int size, ICacheKey key, out BufferHolder holder)
{
return _cachedConvertedBuffers.TryGetValue(offset, size, key, out holder);
}
public void AddCachedConvertedBuffer(int offset, int size, ICacheKey key, BufferHolder holder)
{
_cachedConvertedBuffers.Add(offset, size, key, holder);
}
public void AddCachedConvertedBufferDependency(int offset, int size, ICacheKey key, Dependency dependency)
{
_cachedConvertedBuffers.AddDependency(offset, size, key, dependency);
}
public void RemoveCachedConvertedBuffer(int offset, int size, ICacheKey key)
{
_cachedConvertedBuffers.Remove(offset, size, key);
}
public void Dispose()
{
_swapQueued = false;
_gd.PipelineInternal?.FlushCommandsIfWeightExceeding(_buffer, (ulong)Size);
_buffer.Dispose();
_allocationAuto.Dispose();
_cachedConvertedBuffers.Dispose();
_flushLock.AcquireWriterLock(Timeout.Infinite);
ClearFlushFence();
_flushLock.ReleaseWriterLock();
}
}
}