mirror of
https://github.com/Ryujinx/Ryujinx.git
synced 2024-08-18 06:38:42 +02:00
Calculate vertex buffer sizes from index buffer (#1663)
* Calculate vertex buffer size from maximum index buffer index * Increase maximum index buffer count for it to be considered profitable for counting
This commit is contained in:
parent
bb8a920b63
commit
10d649e6d3
4 changed files with 352 additions and 5 deletions
307
Ryujinx.Graphics.Gpu/Engine/Threed/IbUtils.cs
Normal file
307
Ryujinx.Graphics.Gpu/Engine/Threed/IbUtils.cs
Normal file
|
@ -0,0 +1,307 @@
|
|||
using Ryujinx.Graphics.GAL;
|
||||
using Ryujinx.Graphics.Gpu.Memory;
|
||||
using System;
|
||||
using System.Runtime.InteropServices;
|
||||
using System.Runtime.Intrinsics;
|
||||
using System.Runtime.Intrinsics.X86;
|
||||
|
||||
namespace Ryujinx.Graphics.Gpu.Engine.Threed
|
||||
{
|
||||
/// <summary>
|
||||
/// Index buffer utility methods.
|
||||
/// </summary>
|
||||
static class IbUtils
|
||||
{
|
||||
/// <summary>
|
||||
/// Minimum size that the vertex buffer must have, in bytes, to make the index counting profitable.
|
||||
/// </summary>
|
||||
private const ulong MinimumVbSizeThreshold = 0x200000; // 2 MB
|
||||
|
||||
/// <summary>
|
||||
/// Maximum number of indices that the index buffer may have to make the index counting profitable.
|
||||
/// </summary>
|
||||
private const int MaximumIndexCountThreshold = 65536;
|
||||
|
||||
/// <summary>
|
||||
/// Checks if getting the vertex buffer size from the maximum index buffer index is worth it.
|
||||
/// </summary>
|
||||
/// <param name="vbSizeMax">Maximum size that the vertex buffer may possibly have, in bytes</param>
|
||||
/// <param name="indexCount">Total number of indices on the index buffer</param>
|
||||
/// <returns>True if getting the vertex buffer size from the index buffer may yield performance improvements</returns>
|
||||
public static bool IsIbCountingProfitable(ulong vbSizeMax, int indexCount)
|
||||
{
|
||||
return vbSizeMax >= MinimumVbSizeThreshold && indexCount <= MaximumIndexCountThreshold;
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Gets the vertex count of the vertex buffer accessed with the indices from the current index buffer.
|
||||
/// </summary>
|
||||
/// <param name="mm">GPU memory manager</param>
|
||||
/// <param name="type">Index buffer element integer type</param>
|
||||
/// <param name="gpuVa">GPU virtual address of the index buffer</param>
|
||||
/// <param name="firstIndex">Index of the first index buffer element used on the draw</param>
|
||||
/// <param name="indexCount">Number of index buffer elements used on the draw</param>
|
||||
/// <returns>Vertex count</returns>
|
||||
public static ulong GetVertexCount(MemoryManager mm, IndexType type, ulong gpuVa, int firstIndex, int indexCount)
|
||||
{
|
||||
return type switch
|
||||
{
|
||||
IndexType.UShort => CountU16(mm, gpuVa, firstIndex, indexCount),
|
||||
IndexType.UInt => CountU32(mm, gpuVa, firstIndex, indexCount),
|
||||
_ => CountU8(mm, gpuVa, firstIndex, indexCount)
|
||||
};
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Gets the vertex count of the vertex buffer accessed with the indices from the current index buffer, with 8-bit indices.
|
||||
/// </summary>
|
||||
/// <param name="mm">GPU memory manager</param>
|
||||
/// <param name="gpuVa">GPU virtual address of the index buffer</param>
|
||||
/// <param name="firstIndex">Index of the first index buffer element used on the draw</param>
|
||||
/// <param name="indexCount">Number of index buffer elements used on the draw</param>
|
||||
/// <returns>Vertex count</returns>
|
||||
private unsafe static ulong CountU8(MemoryManager mm, ulong gpuVa, int firstIndex, int indexCount)
|
||||
{
|
||||
uint max = 0;
|
||||
ReadOnlySpan<byte> data = mm.GetSpan(gpuVa, firstIndex + indexCount);
|
||||
|
||||
if (Avx2.IsSupported)
|
||||
{
|
||||
fixed (byte* pInput = data)
|
||||
{
|
||||
int endAligned = firstIndex + ((data.Length - firstIndex) & ~127);
|
||||
|
||||
var result = Vector256<byte>.Zero;
|
||||
|
||||
for (int i = firstIndex; i < endAligned; i += 128)
|
||||
{
|
||||
var dataVec0 = Avx.LoadVector256(pInput + (nuint)(uint)i);
|
||||
var dataVec1 = Avx.LoadVector256(pInput + (nuint)(uint)i + 32);
|
||||
var dataVec2 = Avx.LoadVector256(pInput + (nuint)(uint)i + 64);
|
||||
var dataVec3 = Avx.LoadVector256(pInput + (nuint)(uint)i + 96);
|
||||
|
||||
var max01 = Avx2.Max(dataVec0, dataVec1);
|
||||
var max23 = Avx2.Max(dataVec2, dataVec3);
|
||||
var max0123 = Avx2.Max(max01, max23);
|
||||
|
||||
result = Avx2.Max(result, max0123);
|
||||
}
|
||||
|
||||
result = Avx2.Max(result, Avx2.Shuffle(result.AsInt32(), 0xee).AsByte());
|
||||
result = Avx2.Max(result, Avx2.Shuffle(result.AsInt32(), 0x55).AsByte());
|
||||
result = Avx2.Max(result, Avx2.ShuffleLow(result.AsUInt16(), 0x55).AsByte());
|
||||
result = Avx2.Max(result, Avx2.ShiftRightLogical(result.AsUInt16(), 8).AsByte());
|
||||
|
||||
max = Math.Max(result.GetElement(0), result.GetElement(16));
|
||||
|
||||
firstIndex = endAligned;
|
||||
}
|
||||
}
|
||||
else if (Sse2.IsSupported)
|
||||
{
|
||||
fixed (byte* pInput = data)
|
||||
{
|
||||
int endAligned = firstIndex + ((data.Length - firstIndex) & ~63);
|
||||
|
||||
var result = Vector128<byte>.Zero;
|
||||
|
||||
for (int i = firstIndex; i < endAligned; i += 64)
|
||||
{
|
||||
var dataVec0 = Sse2.LoadVector128(pInput + (nuint)(uint)i);
|
||||
var dataVec1 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 16);
|
||||
var dataVec2 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 32);
|
||||
var dataVec3 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 48);
|
||||
|
||||
var max01 = Sse2.Max(dataVec0, dataVec1);
|
||||
var max23 = Sse2.Max(dataVec2, dataVec3);
|
||||
var max0123 = Sse2.Max(max01, max23);
|
||||
|
||||
result = Sse2.Max(result, max0123);
|
||||
}
|
||||
|
||||
result = Sse2.Max(result, Sse2.Shuffle(result.AsInt32(), 0xee).AsByte());
|
||||
result = Sse2.Max(result, Sse2.Shuffle(result.AsInt32(), 0x55).AsByte());
|
||||
result = Sse2.Max(result, Sse2.ShuffleLow(result.AsUInt16(), 0x55).AsByte());
|
||||
result = Sse2.Max(result, Sse2.ShiftRightLogical(result.AsUInt16(), 8).AsByte());
|
||||
|
||||
max = result.GetElement(0);
|
||||
|
||||
firstIndex = endAligned;
|
||||
}
|
||||
}
|
||||
|
||||
for (int i = firstIndex; i < data.Length; i++)
|
||||
{
|
||||
if (max < data[i]) max = data[i];
|
||||
}
|
||||
|
||||
return (ulong)max + 1;
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Gets the vertex count of the vertex buffer accessed with the indices from the current index buffer, with 16-bit indices.
|
||||
/// </summary>
|
||||
/// <param name="mm">GPU memory manager</param>
|
||||
/// <param name="gpuVa">GPU virtual address of the index buffer</param>
|
||||
/// <param name="firstIndex">Index of the first index buffer element used on the draw</param>
|
||||
/// <param name="indexCount">Number of index buffer elements used on the draw</param>
|
||||
/// <returns>Vertex count</returns>
|
||||
private unsafe static ulong CountU16(MemoryManager mm, ulong gpuVa, int firstIndex, int indexCount)
|
||||
{
|
||||
uint max = 0;
|
||||
ReadOnlySpan<ushort> data = MemoryMarshal.Cast<byte, ushort>(mm.GetSpan(gpuVa, (firstIndex + indexCount) * 2));
|
||||
|
||||
if (Avx2.IsSupported)
|
||||
{
|
||||
fixed (ushort* pInput = data)
|
||||
{
|
||||
int endAligned = firstIndex + ((data.Length - firstIndex) & ~63);
|
||||
|
||||
var result = Vector256<ushort>.Zero;
|
||||
|
||||
for (int i = firstIndex; i < endAligned; i += 64)
|
||||
{
|
||||
var dataVec0 = Avx.LoadVector256(pInput + (nuint)(uint)i);
|
||||
var dataVec1 = Avx.LoadVector256(pInput + (nuint)(uint)i + 16);
|
||||
var dataVec2 = Avx.LoadVector256(pInput + (nuint)(uint)i + 32);
|
||||
var dataVec3 = Avx.LoadVector256(pInput + (nuint)(uint)i + 48);
|
||||
|
||||
var max01 = Avx2.Max(dataVec0, dataVec1);
|
||||
var max23 = Avx2.Max(dataVec2, dataVec3);
|
||||
var max0123 = Avx2.Max(max01, max23);
|
||||
|
||||
result = Avx2.Max(result, max0123);
|
||||
}
|
||||
|
||||
result = Avx2.Max(result, Avx2.Shuffle(result.AsInt32(), 0xee).AsUInt16());
|
||||
result = Avx2.Max(result, Avx2.Shuffle(result.AsInt32(), 0x55).AsUInt16());
|
||||
result = Avx2.Max(result, Avx2.ShuffleLow(result, 0x55));
|
||||
|
||||
max = Math.Max(result.GetElement(0), result.GetElement(8));
|
||||
|
||||
firstIndex = endAligned;
|
||||
}
|
||||
}
|
||||
else if (Sse41.IsSupported)
|
||||
{
|
||||
fixed (ushort* pInput = data)
|
||||
{
|
||||
int endAligned = firstIndex + ((data.Length - firstIndex) & ~31);
|
||||
|
||||
var result = Vector128<ushort>.Zero;
|
||||
|
||||
for (int i = firstIndex; i < endAligned; i += 32)
|
||||
{
|
||||
var dataVec0 = Sse2.LoadVector128(pInput + (nuint)(uint)i);
|
||||
var dataVec1 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 8);
|
||||
var dataVec2 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 16);
|
||||
var dataVec3 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 24);
|
||||
|
||||
var max01 = Sse41.Max(dataVec0, dataVec1);
|
||||
var max23 = Sse41.Max(dataVec2, dataVec3);
|
||||
var max0123 = Sse41.Max(max01, max23);
|
||||
|
||||
result = Sse41.Max(result, max0123);
|
||||
}
|
||||
|
||||
result = Sse41.Max(result, Sse2.Shuffle(result.AsInt32(), 0xee).AsUInt16());
|
||||
result = Sse41.Max(result, Sse2.Shuffle(result.AsInt32(), 0x55).AsUInt16());
|
||||
result = Sse41.Max(result, Sse2.ShuffleLow(result, 0x55));
|
||||
|
||||
max = result.GetElement(0);
|
||||
|
||||
firstIndex = endAligned;
|
||||
}
|
||||
}
|
||||
|
||||
for (int i = firstIndex; i < data.Length; i++)
|
||||
{
|
||||
if (max < data[i]) max = data[i];
|
||||
}
|
||||
|
||||
return (ulong)max + 1;
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Gets the vertex count of the vertex buffer accessed with the indices from the current index buffer, with 32-bit indices.
|
||||
/// </summary>
|
||||
/// <param name="mm">GPU memory manager</param>
|
||||
/// <param name="gpuVa">GPU virtual address of the index buffer</param>
|
||||
/// <param name="firstIndex">Index of the first index buffer element used on the draw</param>
|
||||
/// <param name="indexCount">Number of index buffer elements used on the draw</param>
|
||||
/// <returns>Vertex count</returns>
|
||||
private unsafe static ulong CountU32(MemoryManager mm, ulong gpuVa, int firstIndex, int indexCount)
|
||||
{
|
||||
uint max = 0;
|
||||
ReadOnlySpan<uint> data = MemoryMarshal.Cast<byte, uint>(mm.GetSpan(gpuVa, (firstIndex + indexCount) * 4));
|
||||
|
||||
if (Avx2.IsSupported)
|
||||
{
|
||||
fixed (uint* pInput = data)
|
||||
{
|
||||
int endAligned = firstIndex + ((data.Length - firstIndex) & ~31);
|
||||
|
||||
var result = Vector256<uint>.Zero;
|
||||
|
||||
for (int i = firstIndex; i < endAligned; i += 32)
|
||||
{
|
||||
var dataVec0 = Avx.LoadVector256(pInput + (nuint)(uint)i);
|
||||
var dataVec1 = Avx.LoadVector256(pInput + (nuint)(uint)i + 8);
|
||||
var dataVec2 = Avx.LoadVector256(pInput + (nuint)(uint)i + 16);
|
||||
var dataVec3 = Avx.LoadVector256(pInput + (nuint)(uint)i + 24);
|
||||
|
||||
var max01 = Avx2.Max(dataVec0, dataVec1);
|
||||
var max23 = Avx2.Max(dataVec2, dataVec3);
|
||||
var max0123 = Avx2.Max(max01, max23);
|
||||
|
||||
result = Avx2.Max(result, max0123);
|
||||
}
|
||||
|
||||
result = Avx2.Max(result, Avx2.Shuffle(result, 0xee));
|
||||
result = Avx2.Max(result, Avx2.Shuffle(result, 0x55));
|
||||
|
||||
max = Math.Max(result.GetElement(0), result.GetElement(4));
|
||||
|
||||
firstIndex = endAligned;
|
||||
}
|
||||
}
|
||||
else if (Sse41.IsSupported)
|
||||
{
|
||||
fixed (uint* pInput = data)
|
||||
{
|
||||
int endAligned = firstIndex + ((data.Length - firstIndex) & ~15);
|
||||
|
||||
var result = Vector128<uint>.Zero;
|
||||
|
||||
for (int i = firstIndex; i < endAligned; i += 16)
|
||||
{
|
||||
var dataVec0 = Sse2.LoadVector128(pInput + (nuint)(uint)i);
|
||||
var dataVec1 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 4);
|
||||
var dataVec2 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 8);
|
||||
var dataVec3 = Sse2.LoadVector128(pInput + (nuint)(uint)i + 12);
|
||||
|
||||
var max01 = Sse41.Max(dataVec0, dataVec1);
|
||||
var max23 = Sse41.Max(dataVec2, dataVec3);
|
||||
var max0123 = Sse41.Max(max01, max23);
|
||||
|
||||
result = Sse41.Max(result, max0123);
|
||||
}
|
||||
|
||||
result = Sse41.Max(result, Sse2.Shuffle(result, 0xee));
|
||||
result = Sse41.Max(result, Sse2.Shuffle(result, 0x55));
|
||||
|
||||
max = result.GetElement(0);
|
||||
|
||||
firstIndex = endAligned;
|
||||
}
|
||||
}
|
||||
|
||||
for (int i = firstIndex; i < data.Length; i++)
|
||||
{
|
||||
if (max < data[i]) max = data[i];
|
||||
}
|
||||
|
||||
return (ulong)max + 1;
|
||||
}
|
||||
}
|
||||
}
|
|
@ -34,6 +34,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
|
|||
private byte _vsClipDistancesWritten;
|
||||
|
||||
private bool _prevDrawIndexed;
|
||||
private int _prevFirstIndex;
|
||||
private int _prevIndexCount;
|
||||
private bool _prevTfEnable;
|
||||
|
||||
/// <summary>
|
||||
|
@ -182,10 +184,25 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
|
|||
// method when doing indexed draws, so we need to make sure
|
||||
// to update the vertex buffers if we are doing a regular
|
||||
// draw after a indexed one and vice-versa.
|
||||
if (_drawState.DrawIndexed != _prevDrawIndexed)
|
||||
if (GraphicsConfig.EnableIndexedVbSizeDetection)
|
||||
{
|
||||
_updateTracker.ForceDirty(VertexBufferStateIndex);
|
||||
_prevDrawIndexed = _drawState.DrawIndexed;
|
||||
if (_drawState.DrawIndexed != _prevDrawIndexed ||
|
||||
_drawState.FirstIndex != _prevFirstIndex ||
|
||||
_drawState.IndexCount != _prevIndexCount)
|
||||
{
|
||||
_updateTracker.ForceDirty(VertexBufferStateIndex);
|
||||
_prevDrawIndexed = _drawState.DrawIndexed;
|
||||
_prevFirstIndex = _drawState.FirstIndex;
|
||||
_prevIndexCount = _drawState.IndexCount;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
if (_drawState.DrawIndexed != _prevDrawIndexed)
|
||||
{
|
||||
_updateTracker.ForceDirty(VertexBufferStateIndex);
|
||||
_prevDrawIndexed = _drawState.DrawIndexed;
|
||||
}
|
||||
}
|
||||
|
||||
bool tfEnable = _state.State.TfEnable;
|
||||
|
@ -782,7 +799,25 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
|
|||
{
|
||||
// This size may be (much) larger than the real vertex buffer size.
|
||||
// Avoid calculating it this way, unless we don't have any other option.
|
||||
size = endAddress.Pack() - address + 1;
|
||||
ulong vbSizeMax = endAddress.Pack() - address + 1;
|
||||
|
||||
int firstIndex = _drawState.FirstIndex;
|
||||
int indexCount = _drawState.IndexCount;
|
||||
|
||||
bool ibCountingProfitable = GraphicsConfig.EnableIndexedVbSizeDetection && IbUtils.IsIbCountingProfitable(vbSizeMax, indexCount);
|
||||
|
||||
if (ibCountingProfitable && !_drawState.IbStreamer.HasInlineIndexData && _drawState.DrawIndexed && stride != 0)
|
||||
{
|
||||
IndexType ibType = _state.State.IndexBufferState.Type;
|
||||
ulong ibGpuVa = _state.State.IndexBufferState.Address.Pack();
|
||||
ulong vertexCount = IbUtils.GetVertexCount(_channel.MemoryManager, ibType, ibGpuVa, firstIndex, indexCount);
|
||||
|
||||
size = Math.Min(vertexCount * (ulong)stride, vbSizeMax);
|
||||
}
|
||||
else
|
||||
{
|
||||
size = vbSizeMax;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
|
|
|
@ -33,6 +33,11 @@ namespace Ryujinx.Graphics.Gpu
|
|||
/// </summary>
|
||||
public static bool EnableMacroJit = true;
|
||||
|
||||
/// <summary>
|
||||
/// Enables or disables vertex buffer size detection from the index buffer, for indexed draws.
|
||||
/// </summary>
|
||||
public static bool EnableIndexedVbSizeDetection = true;
|
||||
|
||||
/// <summary>
|
||||
/// Title id of the current running game.
|
||||
/// Used by the shader cache.
|
||||
|
|
Loading…
Reference in a new issue