rjx-mirror/Ryujinx.Graphics.Vulkan/PipelineFull.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

314 lines
10 KiB
C#

using Ryujinx.Graphics.GAL;
using Ryujinx.Graphics.Vulkan.Queries;
using Silk.NET.Vulkan;
using System;
using System.Collections.Generic;
namespace Ryujinx.Graphics.Vulkan
{
class PipelineFull : PipelineBase, IPipeline
{
private const ulong MinByteWeightForFlush = 256 * 1024 * 1024; // MiB
private readonly List<(QueryPool, bool)> _activeQueries;
private CounterQueueEvent _activeConditionalRender;
private readonly List<BufferedQuery> _pendingQueryCopies;
private ulong _byteWeight;
private List<BufferHolder> _backingSwaps;
public PipelineFull(VulkanRenderer gd, Device device) : base(gd, device)
{
_activeQueries = new List<(QueryPool, bool)>();
_pendingQueryCopies = new();
_backingSwaps = new();
CommandBuffer = (Cbs = gd.CommandBufferPool.Rent()).CommandBuffer;
}
private void CopyPendingQuery()
{
foreach (var query in _pendingQueryCopies)
{
query.PoolCopy(Cbs);
}
_pendingQueryCopies.Clear();
}
public void ClearRenderTargetColor(int index, int layer, int layerCount, uint componentMask, ColorF color)
{
if (FramebufferParams == null)
{
return;
}
if (componentMask != 0xf)
{
// We can't use CmdClearAttachments if not writing all components,
// because on Vulkan, the pipeline state does not affect clears.
var dstTexture = FramebufferParams.GetAttachment(index);
if (dstTexture == null)
{
return;
}
Span<float> clearColor = stackalloc float[4];
clearColor[0] = color.Red;
clearColor[1] = color.Green;
clearColor[2] = color.Blue;
clearColor[3] = color.Alpha;
// TODO: Clear only the specified layer.
Gd.HelperShader.Clear(
Gd,
dstTexture,
clearColor,
componentMask,
(int)FramebufferParams.Width,
(int)FramebufferParams.Height,
FramebufferParams.AttachmentFormats[index],
FramebufferParams.GetAttachmentComponentType(index),
ClearScissor);
}
else
{
ClearRenderTargetColor(index, layer, layerCount, color);
}
}
public void EndHostConditionalRendering()
{
if (Gd.Capabilities.SupportsConditionalRendering)
{
// Gd.ConditionalRenderingApi.CmdEndConditionalRendering(CommandBuffer);
}
else
{
// throw new NotSupportedException();
}
_activeConditionalRender?.ReleaseHostAccess();
_activeConditionalRender = null;
}
public bool TryHostConditionalRendering(ICounterEvent value, ulong compare, bool isEqual)
{
// Compare an event and a constant value.
if (value is CounterQueueEvent evt)
{
// Easy host conditional rendering when the check matches what GL can do:
// - Event is of type samples passed.
// - Result is not a combination of multiple queries.
// - Comparing against 0.
// - Event has not already been flushed.
if (compare == 0 && evt.Type == CounterType.SamplesPassed && evt.ClearCounter)
{
if (!value.ReserveForHostAccess())
{
// If the event has been flushed, then just use the values on the CPU.
// The query object may already be repurposed for another draw (eg. begin + end).
return false;
}
if (Gd.Capabilities.SupportsConditionalRendering)
{
var buffer = evt.GetBuffer().Get(Cbs, 0, sizeof(long)).Value;
var flags = isEqual ? ConditionalRenderingFlagsEXT.InvertedBitExt : 0;
var conditionalRenderingBeginInfo = new ConditionalRenderingBeginInfoEXT()
{
SType = StructureType.ConditionalRenderingBeginInfoExt,
Buffer = buffer,
Flags = flags
};
// Gd.ConditionalRenderingApi.CmdBeginConditionalRendering(CommandBuffer, conditionalRenderingBeginInfo);
}
_activeConditionalRender = evt;
return true;
}
}
// The GPU will flush the queries to CPU and evaluate the condition there instead.
FlushPendingQuery(); // The thread will be stalled manually flushing the counter, so flush commands now.
return false;
}
public bool TryHostConditionalRendering(ICounterEvent value, ICounterEvent compare, bool isEqual)
{
FlushPendingQuery(); // The thread will be stalled manually flushing the counter, so flush commands now.
return false;
}
private void FlushPendingQuery()
{
if (AutoFlush.ShouldFlushQuery())
{
FlushCommandsImpl();
}
}
public CommandBufferScoped GetPreloadCommandBuffer()
{
if (PreloadCbs == null)
{
PreloadCbs = Gd.CommandBufferPool.Rent();
}
return PreloadCbs.Value;
}
public void FlushCommandsIfWeightExceeding(IAuto disposedResource, ulong byteWeight)
{
bool usedByCurrentCb = disposedResource.HasCommandBufferDependency(Cbs);
if (PreloadCbs != null && !usedByCurrentCb)
{
usedByCurrentCb = disposedResource.HasCommandBufferDependency(PreloadCbs.Value);
}
if (usedByCurrentCb)
{
// Since we can only free memory after the command buffer that uses a given resource was executed,
// keeping the command buffer might cause a high amount of memory to be in use.
// To prevent that, we force submit command buffers if the memory usage by resources
// in use by the current command buffer is above a given limit, and those resources were disposed.
_byteWeight += byteWeight;
if (_byteWeight >= MinByteWeightForFlush)
{
FlushCommandsImpl();
}
}
}
private void TryBackingSwaps()
{
CommandBufferScoped? cbs = null;
_backingSwaps.RemoveAll((holder) => holder.TryBackingSwap(ref cbs));
cbs?.Dispose();
}
public void AddBackingSwap(BufferHolder holder)
{
_backingSwaps.Add(holder);
}
public void Restore()
{
if (Pipeline != null)
{
Gd.Api.CmdBindPipeline(CommandBuffer, Pbp, Pipeline.Get(Cbs).Value);
}
SignalCommandBufferChange();
DynamicState.ReplayIfDirty(Gd.Api, CommandBuffer);
}
public void FlushCommandsImpl()
{
AutoFlush.RegisterFlush(DrawCount);
EndRenderPass();
foreach ((var queryPool, _) in _activeQueries)
{
Gd.Api.CmdEndQuery(CommandBuffer, queryPool, 0);
}
_byteWeight = 0;
if (PreloadCbs != null)
{
PreloadCbs.Value.Dispose();
PreloadCbs = null;
}
CommandBuffer = (Cbs = Gd.CommandBufferPool.ReturnAndRent(Cbs)).CommandBuffer;
Gd.RegisterFlush();
// Restore per-command buffer state.
foreach ((var queryPool, var isOcclusion) in _activeQueries)
{
bool isPrecise = Gd.Capabilities.SupportsPreciseOcclusionQueries && isOcclusion;
Gd.Api.CmdResetQueryPool(CommandBuffer, queryPool, 0, 1);
Gd.Api.CmdBeginQuery(CommandBuffer, queryPool, 0, isPrecise ? QueryControlFlags.PreciseBit : 0);
}
Gd.ResetCounterPool();
TryBackingSwaps();
Restore();
}
public void BeginQuery(BufferedQuery query, QueryPool pool, bool needsReset, bool isOcclusion, bool fromSamplePool)
{
if (needsReset)
{
EndRenderPass();
Gd.Api.CmdResetQueryPool(CommandBuffer, pool, 0, 1);
if (fromSamplePool)
{
// Try reset some additional queries in advance.
Gd.ResetFutureCounters(CommandBuffer, AutoFlush.GetRemainingQueries());
}
}
bool isPrecise = Gd.Capabilities.SupportsPreciseOcclusionQueries && isOcclusion;
Gd.Api.CmdBeginQuery(CommandBuffer, pool, 0, isPrecise ? QueryControlFlags.PreciseBit : 0);
_activeQueries.Add((pool, isOcclusion));
}
public void EndQuery(QueryPool pool)
{
Gd.Api.CmdEndQuery(CommandBuffer, pool, 0);
for (int i = 0; i < _activeQueries.Count; i++)
{
if (_activeQueries[i].Item1.Handle == pool.Handle)
{
_activeQueries.RemoveAt(i);
break;
}
}
}
public void CopyQueryResults(BufferedQuery query)
{
_pendingQueryCopies.Add(query);
if (AutoFlush.RegisterPendingQuery())
{
FlushCommandsImpl();
}
}
protected override void SignalAttachmentChange()
{
if (AutoFlush.ShouldFlushAttachmentChange(DrawCount))
{
FlushCommandsImpl();
}
}
protected override void SignalRenderPassEnd()
{
CopyPendingQuery();
}
}
}