Enable multithreaded VP9 decoding (#2009)

* Enable multithreaded VP9 decoding

* Limit the number of threads used for video decoding
This commit is contained in:
gdkchan 2021-02-10 21:54:42 -03:00 committed by GitHub
parent 172ec326e5
commit c465d771dd
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 252 additions and 9 deletions

View File

@ -1,13 +1,14 @@
using Ryujinx.Common.Memory;
using Ryujinx.Graphics.Nvdec.Vp9.Common;
using Ryujinx.Graphics.Nvdec.Vp9.Dsp;
using Ryujinx.Graphics.Nvdec.Vp9.Types;
using Ryujinx.Graphics.Video;
using System;
using System.Buffers.Binary;
using System.Diagnostics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using Ryujinx.Graphics.Nvdec.Vp9.Common;
using Ryujinx.Graphics.Nvdec.Vp9.Dsp;
using Ryujinx.Graphics.Nvdec.Vp9.Types;
using Ryujinx.Graphics.Video;
using System.Threading.Tasks;
using Mv = Ryujinx.Graphics.Nvdec.Vp9.Types.Mv;
namespace Ryujinx.Graphics.Nvdec.Vp9
@ -1095,6 +1096,19 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
data = data.Slice(size);
}
private static void GetTileBuffers(ref Vp9Common cm, ArrayPtr<byte> data, int tileCols, ref Array64<TileBuffer> tileBuffers)
{
int c;
for (c = 0; c < tileCols; ++c)
{
bool isLast = c == tileCols - 1;
ref TileBuffer buf = ref tileBuffers[c];
buf.Col = c;
GetTileBuffer(isLast, ref cm.Error, ref data, ref buf);
}
}
private static void GetTileBuffers(
ref Vp9Common cm,
ArrayPtr<byte> data,
@ -1181,5 +1195,163 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
// Get last tile data.
return cm.TileWorkerData[tileCols * tileRows - 1].BitReader.FindEnd();
}
private static bool DecodeTileCol(ref TileWorkerData tileData, ref Vp9Common cm, ref Array64<TileBuffer> tileBuffers)
{
ref TileInfo tile = ref tileData.Xd.Tile;
int finalCol = (1 << cm.Log2TileCols) - 1;
ArrayPtr<byte> bitReaderEnd = ArrayPtr<byte>.Null;
int n = tileData.BufStart;
tileData.Xd.Corrupted = false;
do
{
ref TileBuffer buf = ref tileBuffers[n];
Debug.Assert(cm.Log2TileRows == 0);
tileData.Dqcoeff = new Array32<Array32<int>>();
tile.Init(ref cm, 0, buf.Col);
SetupTokenDecoder(buf.Data, buf.Size, ref tileData.ErrorInfo, ref tileData.BitReader);
cm.InitMacroBlockD(ref tileData.Xd, new ArrayPtr<int>(ref tileData.Dqcoeff[0][0], 32 * 32));
tileData.Xd.ErrorInfo = new Ptr<InternalErrorInfo>(ref tileData.ErrorInfo);
for (int miRow = tile.MiRowStart; miRow < tile.MiRowEnd; miRow += Constants.MiBlockSize)
{
tileData.Xd.LeftContext = new Array3<Array16<sbyte>>();
tileData.Xd.LeftSegContext = new Array8<sbyte>();
for (int miCol = tile.MiColStart; miCol < tile.MiColEnd; miCol += Constants.MiBlockSize)
{
DecodePartition(ref tileData, ref cm, miRow, miCol, BlockSize.Block64x64, 4);
}
}
if (buf.Col == finalCol)
{
bitReaderEnd = tileData.BitReader.FindEnd();
}
} while (!tileData.Xd.Corrupted && ++n <= tileData.BufEnd);
tileData.DataEnd = bitReaderEnd;
return !tileData.Xd.Corrupted;
}
public static unsafe ArrayPtr<byte> DecodeTilesMt(ref Vp9Common cm, ArrayPtr<byte> data, int maxThreads)
{
ArrayPtr<byte> bitReaderEnd = ArrayPtr<byte>.Null;
int tileCols = 1 << cm.Log2TileCols;
int tileRows = 1 << cm.Log2TileRows;
int totalTiles = tileCols * tileRows;
int numWorkers = Math.Min(maxThreads, tileCols);
int n;
Debug.Assert(tileCols <= (1 << 6));
Debug.Assert(tileRows == 1);
cm.AboveContext.ToSpan().Fill(0);
cm.AboveSegContext.ToSpan().Fill(0);
for (n = 0; n < numWorkers; ++n)
{
ref TileWorkerData tileData = ref cm.TileWorkerData[n + totalTiles];
tileData.Xd = cm.Mb;
tileData.Xd.Counts = new Ptr<Vp9BackwardUpdates>(ref tileData.Counts);
tileData.Counts = new Vp9BackwardUpdates();
}
Array64<TileBuffer> tileBuffers = new Array64<TileBuffer>();
GetTileBuffers(ref cm, data, tileCols, ref tileBuffers);
tileBuffers.ToSpan().Slice(0, tileCols).Sort(CompareTileBuffers);
if (numWorkers == tileCols)
{
TileBuffer largest = tileBuffers[0];
Span<TileBuffer> buffers = tileBuffers.ToSpan();
buffers.Slice(1).CopyTo(buffers.Slice(0, tileBuffers.Length - 1));
tileBuffers[tileCols - 1] = largest;
}
else
{
int start = 0, end = tileCols - 2;
TileBuffer tmp;
// Interleave the tiles to distribute the load between threads, assuming a
// larger tile implies it is more difficult to decode.
while (start < end)
{
tmp = tileBuffers[start];
tileBuffers[start] = tileBuffers[end];
tileBuffers[end] = tmp;
start += 2;
end -= 2;
}
}
int baseVal = tileCols / numWorkers;
int remain = tileCols % numWorkers;
int bufStart = 0;
for (n = 0; n < numWorkers; ++n)
{
int count = baseVal + (remain + n) / numWorkers;
ref TileWorkerData tileData = ref cm.TileWorkerData[n + totalTiles];
tileData.BufStart = bufStart;
tileData.BufEnd = bufStart + count - 1;
tileData.DataEnd = data.Slice(data.Length);
bufStart += count;
}
Ptr<Vp9Common> cmPtr = new Ptr<Vp9Common>(ref cm);
Parallel.For(0, numWorkers, (n) =>
{
ref TileWorkerData tileData = ref cmPtr.Value.TileWorkerData[n + totalTiles];
if (!DecodeTileCol(ref tileData, ref cmPtr.Value, ref tileBuffers))
{
cmPtr.Value.Mb.Corrupted = true;
}
});
for (; n > 0; --n)
{
if (bitReaderEnd.IsNull)
{
ref TileWorkerData tileData = ref cm.TileWorkerData[n - 1 + totalTiles];
bitReaderEnd = tileData.DataEnd;
}
}
for (n = 0; n < numWorkers; ++n)
{
ref TileWorkerData tileData = ref cm.TileWorkerData[n + totalTiles];
AccumulateFrameCounts(ref cm.Counts.Value, ref tileData.Counts);
}
Debug.Assert(!bitReaderEnd.IsNull || cm.Mb.Corrupted);
return bitReaderEnd;
}
private static int CompareTileBuffers(TileBuffer bufA, TileBuffer bufB)
{
return (bufA.Size < bufB.Size ? 1 : 0) - (bufA.Size > bufB.Size ? 1 : 0);
}
private static void AccumulateFrameCounts(ref Vp9BackwardUpdates accum, ref Vp9BackwardUpdates counts)
{
Span<uint> a = MemoryMarshal.Cast<Vp9BackwardUpdates, uint>(MemoryMarshal.CreateSpan(ref accum, 1));
Span<uint> c = MemoryMarshal.Cast<Vp9BackwardUpdates, uint>(MemoryMarshal.CreateSpan(ref counts, 1));
for (int i = 0; i < a.Length; i++)
{
a[i] += c[i];
}
}
}
}

View File

@ -92,7 +92,14 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
cm.Mb.SetupBlockPlanes(1, 1);
cm.AllocTileWorkerData(_allocator, 1 << pictureInfo.Log2TileCols, 1 << pictureInfo.Log2TileRows);
int tileCols = 1 << pictureInfo.Log2TileCols;
int tileRows = 1 << pictureInfo.Log2TileRows;
// Video usually have only 4 columns, so more threads won't make a difference for those.
// Try to not take all CPU cores for video decoding.
int maxThreads = Math.Min(4, Environment.ProcessorCount / 2);
cm.AllocTileWorkerData(_allocator, tileCols, tileRows, maxThreads);
cm.AllocContextBuffers(_allocator, output.Width, output.Height);
cm.InitContextBuffers();
cm.SetupSegmentationDequant();
@ -104,7 +111,14 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
{
try
{
DecodeFrame.DecodeTiles(ref cm, new ArrayPtr<byte>(dataPtr, bitstream.Length));
if (maxThreads > 1 && tileRows == 1 && tileCols > 1)
{
DecodeFrame.DecodeTilesMt(ref cm, new ArrayPtr<byte>(dataPtr, bitstream.Length), maxThreads);
}
else
{
DecodeFrame.DecodeTiles(ref cm, new ArrayPtr<byte>(dataPtr, bitstream.Length));
}
}
catch (InternalErrorException)
{

View File

@ -87,6 +87,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
return rv;
}
[SkipLocalsInit]
public static void Iwht4x416Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
/* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
@ -142,6 +143,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Iwht4x41Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i;
@ -209,6 +211,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[3] = WrapLow(DctConstRoundShift(s0 + s1 - s3));
}
[SkipLocalsInit]
public static void Idct4(ReadOnlySpan<int> input, Span<int> output)
{
Span<short> step = stackalloc short[4];
@ -231,6 +234,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[3] = WrapLow(step[0] - step[3]);
}
[SkipLocalsInit]
public static void Idct4x416Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -359,6 +363,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[7] = WrapLow(-x1);
}
[SkipLocalsInit]
public static void Idct8(ReadOnlySpan<int> input, Span<int> output)
{
Span<short> step1 = stackalloc short[8];
@ -416,6 +421,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[7] = WrapLow(step1[0] - step1[7]);
}
[SkipLocalsInit]
public static void Idct8x864Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -449,6 +455,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Idct8x812Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -457,6 +464,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[8];
Span<int> tempOut = stackalloc int[8];
output.Fill(0);
// First transform rows
// Only first 4 row has non-zero coefs
for (i = 0; i < 4; ++i)
@ -671,6 +680,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[15] = WrapLow(-x1);
}
[SkipLocalsInit]
public static void Idct16(ReadOnlySpan<int> input, Span<int> output)
{
Span<short> step1 = stackalloc short[16];
@ -838,6 +848,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[15] = WrapLow(step2[0] - step2[15]);
}
[SkipLocalsInit]
public static void Idct16x16256Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -870,6 +881,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Idct16x1638Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -878,6 +890,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[16];
Span<int> tempOut = stackalloc int[16];
output.Fill(0);
// First transform rows. Since all non-zero dct coefficients are in
// upper-left 8x8 area, we only need to calculate first 8 rows here.
for (i = 0; i < 8; ++i)
@ -903,6 +917,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Idct16x1610Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -911,6 +926,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[16];
Span<int> tempOut = stackalloc int[16];
output.Fill(0);
// First transform rows. Since all non-zero dct coefficients are in
// upper-left 4x4 area, we only need to calculate first 4 rows here.
for (i = 0; i < 4; ++i)
@ -955,6 +972,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Idct32(ReadOnlySpan<int> input, Span<int> output)
{
Span<short> step1 = stackalloc short[32];
@ -1324,6 +1342,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[31] = WrapLow(step1[0] - step1[31]);
}
[SkipLocalsInit]
public static void Idct32x321024Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -1370,6 +1389,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Idct32x32135Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -1378,6 +1398,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[32];
Span<int> tempOut = stackalloc int[32];
output.Fill(0);
// Rows
// Only upper-left 16x16 has non-zero coeff
for (i = 0; i < 16; ++i)
@ -1403,6 +1425,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void Idct32x3234Add(ReadOnlySpan<int> input, Span<byte> dest, int stride)
{
int i, j;
@ -1411,6 +1434,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[32];
Span<int> tempOut = stackalloc int[32];
output.Fill(0);
// Rows
// Only upper-left 8x8 has non-zero coeff
for (i = 0; i < 8; ++i)
@ -1456,6 +1481,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIwht4x416Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
/* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
@ -1511,6 +1537,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIwht4x41Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i;
@ -1584,6 +1611,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[3] = HighbdWrapLow(DctConstRoundShift(s0 + s1 - s3), bd);
}
[SkipLocalsInit]
public static void HighbdIdct4(ReadOnlySpan<int> input, Span<int> output, int bd)
{
Span<int> step = stackalloc int[4];
@ -1613,6 +1641,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[3] = HighbdWrapLow(step[0] - step[3], bd);
}
[SkipLocalsInit]
public static void HighbdIdct4x416Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -1748,6 +1777,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[7] = HighbdWrapLow(-x1, bd);
}
[SkipLocalsInit]
public static void HighbdIdct8(ReadOnlySpan<int> input, Span<int> output, int bd)
{
Span<int> step1 = stackalloc int[8];
@ -1803,6 +1833,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[7] = HighbdWrapLow(step1[0] - step1[7], bd);
}
[SkipLocalsInit]
public static void HighbdIdct8x864Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -1835,6 +1866,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIdct8x812Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -1843,6 +1875,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[8];
Span<int> tempOut = stackalloc int[8];
output.Fill(0);
// First transform rows
// Only first 4 row has non-zero coefs
for (i = 0; i < 4; ++i)
@ -2062,6 +2096,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[15] = HighbdWrapLow(-x1, bd);
}
[SkipLocalsInit]
public static void HighbdIdct16(ReadOnlySpan<int> input, Span<int> output, int bd)
{
Span<int> step1 = stackalloc int[16];
@ -2236,6 +2271,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[15] = HighbdWrapLow(step2[0] - step2[15], bd);
}
[SkipLocalsInit]
public static void HighbdIdct16x16256Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -2268,6 +2304,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIdct16x1638Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -2276,6 +2313,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[16];
Span<int> tempOut = stackalloc int[16];
output.Fill(0);
// First transform rows. Since all non-zero dct coefficients are in
// upper-left 8x8 area, we only need to calculate first 8 rows here.
for (i = 0; i < 8; ++i)
@ -2303,6 +2342,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIdct16x1610Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -2311,6 +2351,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[16];
Span<int> tempOut = stackalloc int[16];
output.Fill(0);
// First transform rows. Since all non-zero dct coefficients are in
// upper-left 4x4 area, we only need to calculate first 4 rows here.
for (i = 0; i < 4; ++i)
@ -2355,6 +2397,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIdct32(ReadOnlySpan<int> input, Span<int> output, int bd)
{
Span<int> step1 = stackalloc int[32];
@ -2539,7 +2582,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
step2[8] = step1[8];
step2[15] = step1[15];
temp1 = -step1[9] * (long)CosPi8_64 + step1[14] * (long)CosPi24_64;
temp2 = step1[9] * (long)CosPi24_64 + step1[14] * (long)CosPi8_64;
temp2 = step1[9] * (long)CosPi24_64 + step1[14] * (long)CosPi8_64;
step2[9] = HighbdWrapLow(DctConstRoundShift(temp1), bd);
step2[14] = HighbdWrapLow(DctConstRoundShift(temp2), bd);
temp1 = -step1[10] * (long)CosPi24_64 - step1[13] * (long)CosPi8_64;
@ -2731,6 +2774,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
output[31] = HighbdWrapLow(step1[0] - step1[31], bd);
}
[SkipLocalsInit]
public static void HighbdIdct32x321024Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -2777,6 +2821,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIdct32x32135Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -2785,6 +2830,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[32];
Span<int> tempOut = stackalloc int[32];
output.Fill(0);
// Rows
// Only upper-left 16x16 has non-zero coeff
for (i = 0; i < 16; ++i)
@ -2812,6 +2859,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
[SkipLocalsInit]
public static void HighbdIdct32x3234Add(ReadOnlySpan<int> input, Span<ushort> dest, int stride, int bd)
{
int i, j;
@ -2820,6 +2868,8 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Span<int> tempIn = stackalloc int[32];
Span<int> tempOut = stackalloc int[32];
output.Fill(0);
// Rows
// Only upper-left 8x8 has non-zero coeff
for (i = 0; i < 8; ++i)

View File

@ -4,6 +4,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
{
internal struct TileBuffer
{
public int Col;
public ArrayPtr<byte> Data;
public int Size;
}

View File

@ -1,14 +1,20 @@
using Ryujinx.Common.Memory;
using Ryujinx.Graphics.Nvdec.Vp9.Dsp;
using Ryujinx.Graphics.Nvdec.Vp9.Types;
using Ryujinx.Graphics.Video;
namespace Ryujinx.Graphics.Nvdec.Vp9
{
internal struct TileWorkerData
{
public ArrayPtr<byte> DataEnd;
public int BufStart;
public int BufEnd;
public Reader BitReader;
public Vp9BackwardUpdates Counts;
public MacroBlockD Xd;
/* dqcoeff are shared by all the planes. So planes must be decoded serially */
public Array32<Array32<int>> Dqcoeff;
public InternalErrorInfo ErrorInfo;
}
}

View File

@ -127,9 +127,9 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Types
MBs = MbRows * MbCols;
}
public void AllocTileWorkerData(MemoryAllocator allocator, int tileCols, int tileRows)
public void AllocTileWorkerData(MemoryAllocator allocator, int tileCols, int tileRows, int maxThreads)
{
TileWorkerData = allocator.Allocate<TileWorkerData>(tileCols * tileRows);
TileWorkerData = allocator.Allocate<TileWorkerData>(tileCols * tileRows + (maxThreads > 1 ? maxThreads : 0));
}
public void FreeTileWorkerData(MemoryAllocator allocator)