public void runClassifyVoxel()
{
var gpu = Gpu.Default;
samplePts = ConvertPointsToFloat3(samplePoints);
// allocate memorys
var d_voxelVerts = gpu.Allocate <int>(numVoxels);
float3[] d_samplePts = gpu.Allocate <float3>(samplePts);
//Copy const values
float3 baseP = new float3((float)basePoint.X, (float)basePoint.Y, (float)basePoint.Z);
gpu.Copy(isoValue, constIsovalue);
gpu.Copy(fusion, constFusion);
gpu.Copy(baseP, constBasePoint);
gpu.Copy(voxelSize, constVoxelSize);
gpu.Copy(gridSize, constGridSize);
gpu.Copy(Tables.VertsTable, verticesTable);
gpu.For(0, numVoxels, i => ClassifyVoxel(i, d_samplePts, d_voxelVerts));
voxelVerts = Gpu.CopyToHost(d_voxelVerts);
Gpu.Free(d_samplePts);
Gpu.Free(d_voxelVerts);
}
public void Dispose()
{
Gpu.Free(fromArr);
Gpu.Free(toArr);
Gpu.Free(fromCacheArr);
Gpu.Free(toCacheArr);
}
private async void MergeBuffersGPU(float[] color, float[] normal, float[] p3D, float[] tex)
{
_pixelBufferImg.Clear((byte)0);
var specular = Settings.Specular;
Gpu.Default.For(0, _imgHeight * _imgWidth, i =>
{
var k = i * 3;
ColorPass(_enviromentLight, color, normal, p3D, tex, specular, k);
ArrayMath.Clamp(color, 0, 255, k, 3);
_pixelBufferImg[k + 0] = (byte)(color[k + 0]);
_pixelBufferImg[k + 1] = (byte)(color[k + 1]);
_pixelBufferImg[k + 2] = (byte)(color[k + 2]);
});
Gpu.Free(color);
Gpu.Free(normal);
Gpu.Free(p3D);
Gpu.Free(tex);
await App.Current?.Dispatcher.InvokeAsync(() =>
{
if (Bitmap == null)
{
return;
}
WriteToBitmap(Bitmap, _pixelBufferImg);
});
}
private float[] ComputeDistancesDatasetGPU(Item query)
{
// gpu
Task <float[]> task = Task <float[]> .Factory.StartNew(() =>
{
lock (gpuLock)
{
float[] queryGpu = Gpu.Default.Allocate(query.Descriptor);
//float[] distancesGpu = Gpu.Default.Allocate<float>(gpuDatasetSize);
Gpu.Default.Launch <float[], float[][], float[], Func <float[], float[], float> >
(ComputeDistancesKernel, launchParam, queryGpu, subDatasetGpu, distancesGpu, Item.GetDistanceSQR);
Gpu.Free(queryGpu);
float[] distancesGpuResultTask = Gpu.CopyToHost(distancesGpu);
//Gpu.Free(distancesGpu);
return(distancesGpuResultTask);
}
});
// cpu
//float[] distances = new float[Dataset.Length];
//Parallel.For(gpuDatasetSize, distances.Length, index =>
//{
// distances[index] = Item.GetDistanceSQR(query.Descriptor, Dataset[index].Descriptor);
//});
float[] distancesGpuResult = task.Result;
//Array.Copy(distancesGpuResult, distances, distancesGpuResult.Length);
return(distancesGpuResult);
}
private static int TestGPU()
{
var length = 32_000_000;
var gpu = Gpu.Default;
var a1 = Enumerable.Repeat(1, length).ToArray();
var a2 = Enumerable.Repeat(1, length).ToArray();
var r = new int[length];
int[] arg1 = gpu.Allocate <int>(a1);
int[] arg2 = gpu.Allocate <int>(a2);
int[] result = gpu.Allocate <int>(r);
for (var i = 0; i < 10; i++)
{
var sw = new Stopwatch();
sw.Start();
var s = TestSummGPU(gpu, length, arg1, arg2, result);
sw.Stop();
Console.WriteLine($"GPU1: {sw.Elapsed.Milliseconds} Summ: {s}");
Console.WriteLine();
}
Gpu.Free(arg1);
Gpu.Free(arg2);
Gpu.Free(result);
return(0);
}
// reduce empty voxel and extract active voxels
public void runExtractActiveVoxels()
{
var gpu = Gpu.Default;
// compute the number of active voxels
List <int> index_voxelActiveList = new List <int>();
for (int i = 0; i < voxelVerts.Length; i++)
{
if (voxelVerts[i] > 0)
{
index_voxelActiveList.Add(i);
}
}
// the index of active voxel
index_voxelActive = index_voxelActiveList.ToArray();
num_voxelActive = index_voxelActive.Length;
// the number of vertices in each active voxel
verts_voxelActive = new int[num_voxelActive];
// the number of all vertices
sumVerts = 0;
Parallel.For(0, num_voxelActive, i =>
{
verts_voxelActive[i] = voxelVerts[index_voxelActive[i]];
});
// execute exclusive scan for finding out the indices of result vertices
var op = new Func <int, int, int>((a, b) => { return(a + b); });
Alea.Session session = new Alea.Session(gpu);
int[] d_verts_voxelActive = Gpu.Default.Allocate <int>(verts_voxelActive);
int[] d_voxelVertsScan = Gpu.Default.Allocate <int>(num_voxelActive);
GpuExtension.Scan <int>(session, d_voxelVertsScan, d_verts_voxelActive, 0, op, 0);
var result_Scan = Gpu.CopyToHost(d_voxelVertsScan);
verts_scanIdx = new int[num_voxelActive];
for (int i = 1; i < num_voxelActive; i++)
{
verts_scanIdx[i] = result_Scan[i - 1];
}
try
{
verts_scanIdx[0] = 0;
}
catch (Exception)
{
throw new Exception("No eligible isosurface can be extracted, please change isovalue.");
}
sumVerts = verts_scanIdx.ElementAt(verts_scanIdx.Length - 1) + verts_voxelActive.ElementAt(verts_voxelActive.Length - 1);
Gpu.Free(d_verts_voxelActive);
Gpu.Free(d_voxelVertsScan);
}
public Tuple <int, int> CompareAbsoluteOpt(double[] source, double[] target, double tolerance, double ThreshholdTol)
{
System.Diagnostics.Debug.WriteLine("starting an absolute comparison on GPU");
if (source.Length != target.Length)
{
throw new ArgumentException("The source and target lengths need to match");
}
double epsilon = ThreshholdTol;
double MaxSource = source.Max();
double MaxTarget = target.Max();
double MinDoseEvaluated = (MaxSource * epsilon);
double zero = 0.0;
double lowMultiplier = (1 - tolerance);
double highMultiplier = (1 + tolerance);
int failed = 0;
int isCounted = 0;
Gpu gpu = Gpu.Default;
// filter doses below threshold
// TODO: should failure be -1?
int dimension = source.Length;
double[] sourceOnGPU = gpu.Allocate(source);
double[] targetOnGPU = gpu.Allocate(target);
double[] isCountedArray = gpu.Allocate <double>(dimension);
double[] sourceOnGPULow = gpu.Allocate <double>(dimension);
double[] sourceOnGPUHigh = gpu.Allocate <double>(dimension);
double[] isGTtol = gpu.Allocate <double>(dimension);
gpu.For(0, dimension, i => sourceOnGPU[i] = (sourceOnGPU[i] > epsilon) ? sourceOnGPU[i] : zero);
gpu.For(0, dimension, i => targetOnGPU[i] = (targetOnGPU[i] > epsilon) ? targetOnGPU[i] : zero);
gpu.For(0, dimension, i => sourceOnGPU[i] = (targetOnGPU[i] > epsilon) ? sourceOnGPU[i] : zero);
gpu.For(0, dimension, i => targetOnGPU[i] = (sourceOnGPU[i] > epsilon) ? targetOnGPU[i] : zero);
gpu.For(0, dimension, i => isCountedArray[i] = (sourceOnGPU[i] > zero) ? 1.0 : zero);
gpu.For(0, dimension, i => sourceOnGPULow[i] = lowMultiplier * sourceOnGPU[i]);
gpu.For(0, dimension, i => sourceOnGPUHigh[i] = highMultiplier * sourceOnGPU[i]);
//determine if relative difference is greater than minDoseEvaluated
// stores 1 as GT minDoseEvaluated is true
gpu.For(0, isGTtol.Length,
i => isGTtol[i] = (targetOnGPU[i] < sourceOnGPULow[i] || targetOnGPU[i] > sourceOnGPUHigh[i]) ? 1 : 0);
isCounted = (int)gpu.Sum(isCountedArray);
failed = (int)gpu.Sum(isGTtol);
Gpu.Free(sourceOnGPU);
Gpu.Free(targetOnGPU);
Gpu.Free(sourceOnGPULow);
Gpu.Free(sourceOnGPUHigh);
Gpu.Free(isCountedArray);
Gpu.Free(isGTtol);
System.Diagnostics.Debug.WriteLine("finished an absolute comparison on GPU");
//gpu.Dispose();
return(new Tuple <int, int>(failed, isCounted));
}
private void DisposeDevInputSamples()
{
if (m_DevInputSamples == null)
{
return;
}
Gpu.Free(m_DevInputSamples);
m_DevInputSamples = null;
}
private void DisposeGpuResources()
{
if (m_DevOverdBs == null)
{
return;
}
Gpu.Free(m_DevOverdBs);
m_DevOverdBs = null;
m_SampleCount = 0;
}
public static void RunGpu()
{
var n = GetData(out var x, out var y);
var result = new float[n];
var gpu = Gpu.Default;
var xDevice = gpu.Allocate <float>(n);
var yDevice = gpu.Allocate <float>(n);
var resultDevice = gpu.Allocate <float>(n);
Gpu.Copy(x, xDevice);
Gpu.Copy(y, yDevice);
var lp = new LaunchParam(16, 256);
gpu.Launch(Kernel, lp, resultDevice, xDevice, yDevice);
Gpu.Copy(resultDevice, result);
Gpu.Free(xDevice);
Gpu.Free(yDevice);
Gpu.Free(resultDevice);
}
private static double[,] CosineSimilarityGpu(Gpu gpu, double[][] dataset)
{
int size = dataset.Length * dataset.Length;
var gpuDataset = gpu.Allocate(dataset);
// Allocate directly on gpu.
var gpuDistances = gpu.Allocate <double>(dataset.Length, dataset.Length);
gpu.For(0, size, index =>
{
int i = index / dataset.Length;
int j = index % dataset.Length;
double dotProduct = 0;
double magnitudeOne = 0;
double magnitudeTwo = 0;
for (int k = 0; k < dataset[i].Length; k++)
{
dotProduct += (dataset[i][k] * dataset[j][k]);
magnitudeOne += (dataset[i][k] * dataset[i][k]);
magnitudeTwo += (dataset[j][k] * dataset[j][k]);
}
double distance = Math.Max(0, 1 - (dotProduct / Math.Sqrt(magnitudeOne * magnitudeTwo)));
gpuDistances[i, j] = distance;
});
// Gpu -> Cpu.
var result = new double[dataset.Length, dataset.Length];
Gpu.Copy(gpuDistances, result);
// Release gpu memory.
Gpu.Free(gpuDataset);
Gpu.Free(gpuDistances);
return(result);
}
protected virtual void Free()
{
Gpu.Free(rneuronsArr);
Gpu.Free(resultsArr);
Gpu.Free(inputsArr);
}
public void Free()
{
Gpu.Free(GPUValues);
Gpu.Free(GPUIndices);
}
public void Free()
{
Gpu.Free(GPUArray);
}
public ComputationResult[] Compute(
Problem[] problemsToSolve,
int streamCount,
Action asyncAction = null,
int warpCount = 2)
// cannot be more warps since more memory should be allocated
{
#if (benchmark)
var totalTiming = new Stopwatch();
totalTiming.Start();
var benchmarkTiming = new Stopwatch();
#endif
var gpu = Gpu.Default;
var n = problemsToSolve.First().size;
var power = 1 << n;
var maximumPermissibleWordLength = (n - 1) * (n - 1);
// in order to atomically add to a checking array (designed for queue consistency) one int is used by four threads, so in a reeeeaally pessimistic case 255 is the maximum number of threads (everyone go to the same vertex)
var maximumWarps = gpu.Device.Attributes.MaxThreadsPerBlock / gpu.Device.Attributes.WarpSize;
if (warpCount > maximumWarps)
{
warpCount = maximumWarps;
}
var problemsPerStream = (problemsToSolve.Count() + streamCount - 1) / streamCount;
var problemsPartitioned = Enumerable.Range(0, streamCount)
.Select(i => problemsToSolve.Skip(problemsPerStream * i)
.Take(problemsPerStream)
.ToArray())
.Where(partition => partition.Length > 0)
.ToArray();
streamCount = problemsPartitioned.Length;
var streams = Enumerable.Range(0, streamCount)
.Select(_ => gpu.CreateStream()).ToArray();
var gpuA = problemsPartitioned.Select(problems => gpu.Allocate <int>(problems.Length * n)).ToArray();
var gpuB = problemsPartitioned.Select(problems => gpu.Allocate <int>(problems.Length * n)).ToArray();
var shortestSynchronizingWordLength = problemsPartitioned.Select(problems => gpu.Allocate <int>(problems.Length)).ToArray();
var isSynchronizable = problemsPartitioned.Select(problems => gpu.Allocate <bool>(problems.Length)).ToArray();
gpu.Copy(n, problemSize);
var queueUpperBound = power / 2 + 1;
var launchParameters = new LaunchParam(
new dim3(1, 1, 1),
new dim3(gpu.Device.Attributes.WarpSize * warpCount, 1, 1),
warpCount * (
sizeof(ushort) * queueUpperBound
+ sizeof(ushort) * n * 2
+ sizeof(byte) * power
)
);
var gpuResultsIsSynchronizable = problemsPartitioned
.Select(problems => new bool[problems.Length])
.ToArray();
var gpuResultsShortestSynchronizingWordLength = problemsPartitioned
.Select(problems => new int[problems.Length])
.ToArray();
for (int stream = 0; stream < streamCount; stream++)
{
var problems = problemsPartitioned[stream];
var matrixA = new int[problems.Length * n];
var matrixB = new int[problems.Length * n];
Parallel.For(0, problems.Length, problem =>
{
Array.ConstrainedCopy(problems[problem].stateTransitioningMatrixA, 0, matrixA, problem * n, n);
Array.ConstrainedCopy(problems[problem].stateTransitioningMatrixB, 0, matrixB, problem * n, n);
});
streams[stream].Copy(matrixA, gpuA[stream]);
streams[stream].Copy(matrixB, gpuB[stream]);
streams[stream].Launch(
Kernel,
launchParameters,
gpuA[stream],
gpuB[stream],
isSynchronizable[stream],
shortestSynchronizingWordLength[stream]
);
}
asyncAction?.Invoke();
for (int stream = 0; stream < streamCount; stream++)
{
#if (benchmark)
benchmarkTiming.Start();
#endif
streams[stream].Synchronize();
#if (benchmark)
benchmarkTiming.Stop();
#endif
streams[stream].Copy(isSynchronizable[stream], gpuResultsIsSynchronizable[stream]);
streams[stream].Copy(shortestSynchronizingWordLength[stream], gpuResultsShortestSynchronizingWordLength[stream]);
}
gpu.Synchronize();
#if (benchmark)
#endif
var results = Enumerable.Range(0, streamCount).SelectMany(i => gpuResultsIsSynchronizable[i].Zip(gpuResultsShortestSynchronizingWordLength[i], (isSyncable, shortestWordLength)
=> new ComputationResult()
{
computationType = ComputationType.GPU,
size = problemsToSolve.First().size,
isSynchronizable = isSyncable,
shortestSynchronizingWordLength = shortestWordLength,
algorithmName = GetType().Name
}
).ToArray()
).ToArray();
foreach (var array in gpuA.AsEnumerable <Array>()
.Concat(gpuB)
.Concat(shortestSynchronizingWordLength)
.Concat(isSynchronizable))
{
Gpu.Free(array);
}
foreach (var stream in streams)
{
stream.Dispose();
}
if (results.Any(result => result.isSynchronizable && result.shortestSynchronizingWordLength > maximumPermissibleWordLength))
{
throw new Exception("Cerny conjecture is false");
}
#if (benchmark)
results[0].benchmarkResult = new BenchmarkResult
{
benchmarkedTime = benchmarkTiming.Elapsed,
totalTime = totalTiming.Elapsed
};
#endif
return(results);
}
private static void DoStuff()
{
const float xs = -2.1F;
const float ys = -1.3F;
const int zoom = 1;
const int width = 10240;
const int height = 10240;
const int blackpixel = (byte)0 | ((byte)0 << 8) | ((byte)0 << 16) | ((byte)255 << 24);
var gpubits = gpu.Allocate <int>(width * height);
Console.WriteLine("Started...");
var watch = new Stopwatch();
watch.Start();
const float dx = 2.6F / zoom / width;
const float dy = 2.6F / zoom / height;
gpu.For(0, width * height, (index) => {
//gpu.For(0, width, (x) => {
//for (var y = 0; y < height; y++) {
var y = index / height;
var x = index - (y * height);
var cr = xs + (x * dx);
var ci = ys + (y * dy);
var zr = 0F;
var zi = 0F;
float zrSquare;
float ziSquare;
byte i = 0;
while (i < 255)
{
zrSquare = zr * zr;
ziSquare = zi * zi;
zi = (2 * zi * zr) + ci;
zr = zrSquare - ziSquare + cr;
i++;
if (zrSquare + ziSquare > 4)
{
break;
}
}
if (i == 255)
{
gpubits[index] = blackpixel;
}
else
{
gpubits[index] = (byte)(i * 25 % 256) | ((byte)(i * 3 % 256) << 8) | (((byte)(i % 256)) << 16) | ((byte)255 << 24);
}
});
watch.Stop();
Console.WriteLine($"Elapsed microseconds: {((double)watch.ElapsedTicks) / Stopwatch.Frequency * 1000000}");
var bits = Gpu.CopyToHost(gpubits);
var bitsHandle = GCHandle.Alloc(bits, GCHandleType.Pinned);
var bitmap = new Bitmap(width, height, width * 4, PixelFormat.Format32bppPArgb, bitsHandle.AddrOfPinnedObject());
bitmap.Save("b.png", ImageFormat.Png);
bitmap.Dispose();
bitsHandle.Free();
Gpu.Free(gpubits);
}
// extract isosurface points using GPU
public List <Point3f> runExtractIsoSurfaceGPU()
{
var gpu = Gpu.Default;
// output arguments
float3[] pts = new float3[12 * num_voxelActive];
float3[] d_pts = Gpu.Default.Allocate <float3>(pts);
float3[] d_resultV = Gpu.Default.Allocate <float3>(sumVerts);
float[] d_cubeValues = Gpu.Default.Allocate <float>(8 * num_voxelActive);
// input arguments
float3[] d_samplePts = Gpu.Default.Allocate <float3>(samplePts);
int[] d_verts_scanIdx = Gpu.Default.Allocate <int>(verts_scanIdx);
int[] d_index_voxelActive = Gpu.Default.Allocate <int>(index_voxelActive);
// const values
gpu.Copy(Vertices, constVertices);
gpu.Copy(EdgeDirection, constEdgeDirection);
gpu.Copy(EdgeConnection, constEdgeConnection);
gpu.Copy(Tables.EdgeTable, edgeTable);
gpu.Copy(Tables.TriangleTable_GPU, triangleTable);
float3 baseP = new float3((float)basePoint.X, (float)basePoint.Y, (float)basePoint.Z);
gpu.Copy(baseP, constBasePoint);
gpu.Copy(isoValue, constIsovalue);
gpu.Copy(scale, constScale);
gpu.For(0, num_voxelActive, i =>
{
//计算grid中的位置
int3 gridPos = calcGridPos(d_index_voxelActive[i], constGridSize.Value);
float3 p = new float3();
p.x = constBasePoint.Value.x + gridPos.x * constVoxelSize.Value.x;
p.y = constBasePoint.Value.y + gridPos.y * constVoxelSize.Value.y;
p.z = constBasePoint.Value.z + gridPos.z * constVoxelSize.Value.z;
//输出所有顶点
float3 a0 = p;
float3 a1 = CreateFloat3(constVoxelSize.Value.x + p.x, 0 + p.y, 0 + p.z);
float3 a2 = CreateFloat3(constVoxelSize.Value.x + p.x, constVoxelSize.Value.y + p.y, 0 + p.z);
float3 a3 = CreateFloat3(0 + p.x, constVoxelSize.Value.y + p.y, 0 + p.z);
float3 a4 = CreateFloat3(0 + p.x, 0 + p.y, constVoxelSize.Value.z + p.z);
float3 a5 = CreateFloat3(constVoxelSize.Value.x + p.x, 0 + p.y, constVoxelSize.Value.z + p.z);
float3 a6 = CreateFloat3(constVoxelSize.Value.x + p.x, constVoxelSize.Value.y + p.y, constVoxelSize.Value.z + p.z);
float3 a7 = CreateFloat3(0 + p.x, constVoxelSize.Value.y + p.y, constVoxelSize.Value.z + p.z);
float distance = constVoxelSize.Value.x * constVoxelSize.Value.x +
constVoxelSize.Value.y * constVoxelSize.Value.y + constVoxelSize.Value.z * constVoxelSize.Value.z;
float radius = distance * constFusion.Value;
//Compute cubeValues of 8 vertices
d_cubeValues[i * 8] = ComputeValue(d_samplePts, a0, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 1] = ComputeValue(d_samplePts, a1, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 2] = ComputeValue(d_samplePts, a2, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 3] = ComputeValue(d_samplePts, a3, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 4] = ComputeValue(d_samplePts, a4, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 5] = ComputeValue(d_samplePts, a5, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 6] = ComputeValue(d_samplePts, a6, d_samplePts.Length, constFusion.Value, radius);
d_cubeValues[i * 8 + 7] = ComputeValue(d_samplePts, a7, d_samplePts.Length, constFusion.Value, radius);
//Check each vertex state
int flag = Compact(d_cubeValues[i * 8], constIsovalue.Value);
flag += Compact(d_cubeValues[i * 8 + 1], constIsovalue.Value) * 2;
flag += Compact(d_cubeValues[i * 8 + 2], constIsovalue.Value) * 4;
flag += Compact(d_cubeValues[i * 8 + 3], constIsovalue.Value) * 8;
flag += Compact(d_cubeValues[i * 8 + 4], constIsovalue.Value) * 16;
flag += Compact(d_cubeValues[i * 8 + 5], constIsovalue.Value) * 32;
flag += Compact(d_cubeValues[i * 8 + 6], constIsovalue.Value) * 64;
flag += Compact(d_cubeValues[i * 8 + 7], constIsovalue.Value) * 128;
//find out which edge intersects the isosurface
int EdgeFlag = edgeTable[flag];
//check whether this voxel is crossed by the isosurface
for (int j = 0; j < 12; j++)
{
//check whether an edge have a point
if ((EdgeFlag & (1 << j)) != 0)
{
//compute t values from two end points on each edge
float Offset = GetOffset(d_cubeValues[i * 8 + constEdgeConnection[j * 2 + 0]], d_cubeValues[i * 8 + constEdgeConnection[j * 2 + 1]], constIsovalue.Value);
float3 pt = new float3();
//get positions
pt.x = constBasePoint.Value.x + (gridPos.x + constVertices[constEdgeConnection[j * 2 + 0] * 3 + 0] + Offset * constEdgeDirection[j * 3 + 0]) * constScale.Value;
pt.y = constBasePoint.Value.y + (gridPos.y + constVertices[constEdgeConnection[j * 2 + 0] * 3 + 1] + Offset * constEdgeDirection[j * 3 + 1]) * constScale.Value;
pt.z = constBasePoint.Value.z + (gridPos.z + constVertices[constEdgeConnection[j * 2 + 0] * 3 + 2] + Offset * constEdgeDirection[j * 3 + 2]) * constScale.Value;
d_pts[12 * i + j] = pt;
}
}
int num = 0;
//Find out points from each triangle
for (int Triangle = 0; Triangle < 5; Triangle++)
{
if (triangleTable[flag * 16 + 3 * Triangle] < 0)
{
break;
}
for (int Corner = 0; Corner < 3; Corner++)
{
int Vertex = triangleTable[flag * 16 + 3 * Triangle + Corner];
float3 pd = CreateFloat3(d_pts[12 * i + Vertex].x, d_pts[12 * i + Vertex].y, d_pts[12 * i + Vertex].z);
d_resultV[d_verts_scanIdx[i] + num] = pd;
num++;
}
}
});
resultVerts = Gpu.CopyToHost(d_resultV);
Gpu.Free(d_resultV);
Gpu.Free(d_pts);
Gpu.Free(d_samplePts);
Gpu.Free(d_verts_scanIdx);
Gpu.Free(d_index_voxelActive);
return(ConvertFloat3ToPoint3f(resultVerts));
}
public int ComputeAction(
Problem[] problemsToSolve,
int problemsReadingIndex,
ComputationResult[] computationResults,
int resultsWritingIndex,
int problemCount,
int streamCount,
Action asyncAction = null)
// cannot be more warps since more memory should be allocated
{
#if (benchmark)
var totalTiming = new Stopwatch();
totalTiming.Start();
var benchmarkTiming = new Stopwatch();
#endif
var CernyConjectureFailingIndex = -1;
var gpu = Gpu.Default;
var n = problemsToSolve[problemsReadingIndex].size;
var power = 1 << n;
var maximumPermissibleWordLength = (n - 1) * (n - 1);
// in order to atomically add to a checking array (designed for queue consistency) one int is used by four threads, so in a reeeeaally pessimistic case 255 is the maximum number of threads (everyone go to the same vertex)
// -1 for the already discovered vertex
// at least 2*n+1 threads i.e. 27
var threads = gpu.Device.Attributes.MaxThreadsPerBlock;
var maximumThreads = Math.Min(
gpu.Device.Attributes.MaxThreadsPerBlock,
32 * 2
);
if (threads > maximumThreads)
{
threads = maximumThreads;
}
if (problemCount < streamCount)
{
streamCount = problemCount;
}
var problemsPerStream = (problemCount + streamCount - 1) / streamCount;
var streams = Enumerable.Range(0, streamCount)
.Select(_ => gpu.CreateStream()).ToArray();
gpu.Copy(n, problemSize);
var launchParameters = new LaunchParam(
new dim3(1 << 10, 1, 1),
new dim3(threads, 1, 1),
2 * n * sizeof(ushort)
);
var gpuResultsIsSynchronizable = new bool[streamCount][];
var gpuAs = new int[streamCount][];
var gpuBs = new int[streamCount][];
var isSynchronizable = new bool[streamCount][];
for (int stream = 0; stream < streamCount; stream++)
{
var offset = stream * problemsPerStream;
var localProblemsCount = Math.Min(problemsPerStream, problemCount - offset);
gpuAs[stream] = gpu.Allocate <int>(localProblemsCount * n);
gpuBs[stream] = gpu.Allocate <int>(localProblemsCount * n);
isSynchronizable[stream] = gpu.Allocate <bool>(localProblemsCount);
var matrixA = new int[localProblemsCount * n];
var matrixB = new int[localProblemsCount * n];
Parallel.For(0, localProblemsCount, problem =>
{
Array.ConstrainedCopy(
problemsToSolve[problemsReadingIndex + offset + problem].stateTransitioningMatrixA,
0,
matrixA,
problem * n,
n
);
Array.ConstrainedCopy(
problemsToSolve[problemsReadingIndex + offset + problem].stateTransitioningMatrixB,
0,
matrixB,
problem * n,
n
);
});
streams[stream].Copy(matrixA, gpuAs[stream]);
streams[stream].Copy(matrixB, gpuBs[stream]);
gpuResultsIsSynchronizable[stream] = new bool[localProblemsCount];
streams[stream].Launch(
Kernel,
launchParameters,
gpuAs[stream],
gpuBs[stream],
isSynchronizable[stream]
);
}
asyncAction?.Invoke();
var streamId = 0;
foreach (var stream in streams)
{
#if (benchmark)
benchmarkTiming.Start();
#endif
stream.Synchronize();
#if (benchmark)
benchmarkTiming.Stop();
#endif
stream.Copy(isSynchronizable[streamId], gpuResultsIsSynchronizable[streamId]);
var offset = streamId * problemsPerStream;
var localProblemsCount = Math.Min(problemsPerStream, problemCount - offset);
if (computationResults == null)
{
resultsWritingIndex = 0;
computationResults = new ComputationResult[resultsWritingIndex + problemCount];
}
for (int i = 0; i < localProblemsCount; i++)
{
computationResults[resultsWritingIndex + offset + i].isSynchronizable = gpuResultsIsSynchronizable[streamId][i];
}
streamId++;
}
#if (benchmark)
gpu.Synchronize();
#endif
foreach (var arrays in new IEnumerable <Array>[] { gpuAs, gpuBs, isSynchronizable })
{
foreach (var array in arrays)
{
Gpu.Free(array);
}
}
foreach (var stream in streams)
{
stream.Dispose();
}
var cpu = new SlimCPUSkipper();
var result = cpu.Compute(problemsToSolve, problemsReadingIndex, computationResults, resultsWritingIndex, problemCount, cpu.GetBestParallelism());
#if (benchmark)
computationResults[resultsWritingIndex].benchmarkResult = new BenchmarkResult
{
benchmarkedTime = benchmarkTiming.Elapsed,
totalTime = totalTiming.Elapsed
};
#endif
return(result);
}
public int ComputeManyWithAction(
Problem[] problemsToSolve,
int problemsReadingIndex,
ComputationResult[] computationResults,
int resultsWritingIndex,
int problemCount,
int streamCount,
Action asyncAction = null)
// cannot be more warps since more memory should be allocated
{
#if (benchmark)
var totalTiming = new Stopwatch();
totalTiming.Start();
var benchmarkTiming = new Stopwatch();
#endif
var CernyConjectureFailingIndex = -1;
var gpu = Gpu.Default;
var n = problemsToSolve[problemsReadingIndex].size;
var power = 1 << n;
const int bitSize = 6;
var maximumPermissibleWordLength = (n - 1) * (n - 1);
// in order to atomically add to a checking array (designed for queue consistency) one int is used by four threads, so in a reeeeaally pessimistic case 255 is the maximum number of threads (everyone go to the same vertex)
// -1 for the already discovered vertex, so its -2 in total for both reasons
// at least 2*n+1 threads i.e. 27
var threads = gpu.Device.Attributes.MaxThreadsPerBlock;
var maximumThreads = Math.Min(
gpu.Device.Attributes.MaxThreadsPerBlock,
((1 << bitSize) - 1) - 1
);
var minimumThreads = n + 1;
if (threads > maximumThreads)
{
threads = maximumThreads;
}
if (threads < minimumThreads)
{
threads = minimumThreads;
}
if (threads > maximumThreads)
{
throw new Exception("Impossible to satisfy");
}
if (problemCount < streamCount)
{
streamCount = problemCount;
}
var problemsPerStream = (problemCount + streamCount - 1) / streamCount;
var streams = Enumerable.Range(0, streamCount)
.Select(_ => gpu.CreateStream()).ToArray();
gpu.Copy(n, problemSize);
var isDiscoveredComplexOffset = (power * sizeof(int) + 8 * sizeof(int) / bitSize - 1) / (8 * sizeof(int) / bitSize);
var launchParameters = new LaunchParam(
new dim3(1 << 9, 1, 1),
new dim3(threads, 1, 1),
sizeof(int) * 3
+ isDiscoveredComplexOffset + (((isDiscoveredComplexOffset % sizeof(int)) & 1) == 1 ? 1 : 0)
+ (power / 2 + 1) * sizeof(ushort) * 2
+ (n + 1) * sizeof(uint)
+ sizeof(bool)
);
var gpuResultsIsSynchronizable = new bool[streamCount][];
var gpuResultsShortestSynchronizingWordLength = new int[streamCount][];
var gpuAs = new int[streamCount][];
var gpuBs = new int[streamCount][];
var shortestSynchronizingWordLength = new int[streamCount][];
var isSynchronizable = new bool[streamCount][];
for (int stream = 0; stream < streamCount; stream++)
{
var offset = stream * problemsPerStream;
var localProblemsCount = Math.Min(problemsPerStream, problemCount - offset);
gpuAs[stream] = gpu.Allocate <int>(localProblemsCount * n);
gpuBs[stream] = gpu.Allocate <int>(localProblemsCount * n);
shortestSynchronizingWordLength[stream] = gpu.Allocate <int>(localProblemsCount);
isSynchronizable[stream] = gpu.Allocate <bool>(localProblemsCount);
var matrixA = new int[localProblemsCount * n];
var matrixB = new int[localProblemsCount * n];
Parallel.For(0, localProblemsCount, problem =>
{
Array.ConstrainedCopy(
problemsToSolve[problemsReadingIndex + offset + problem].stateTransitioningMatrixA,
0,
matrixA,
problem * n,
n
);
Array.ConstrainedCopy(
problemsToSolve[problemsReadingIndex + offset + problem].stateTransitioningMatrixB,
0,
matrixB,
problem * n,
n
);
});
streams[stream].Copy(matrixA, gpuAs[stream]);
streams[stream].Copy(matrixB, gpuBs[stream]);
gpuResultsIsSynchronizable[stream] = new bool[localProblemsCount];
gpuResultsShortestSynchronizingWordLength[stream] = new int[localProblemsCount];
streams[stream].Launch(
Kernel,
launchParameters,
gpuAs[stream],
gpuBs[stream],
isSynchronizable[stream],
shortestSynchronizingWordLength[stream]
);
}
asyncAction?.Invoke();
var streamId = 0;
foreach (var stream in streams)
{
#if (benchmark)
benchmarkTiming.Start();
#endif
stream.Synchronize();
#if (benchmark)
benchmarkTiming.Stop();
#endif
stream.Copy(isSynchronizable[streamId], gpuResultsIsSynchronizable[streamId]);
stream.Copy(shortestSynchronizingWordLength[streamId], gpuResultsShortestSynchronizingWordLength[streamId]);
streamId++;
}
#if (benchmark)
gpu.Synchronize();
#endif
Parallel.For(0, streamCount, stream =>
{
var offset = stream * problemsPerStream;
var localProblemsCount = Math.Min(problemsPerStream, problemCount - offset);
for (int i = 0; i < localProblemsCount; i++)
{
if (computationResults != null)
{
computationResults[resultsWritingIndex + offset + i] = new ComputationResult()
{
computationType = ComputationType.GPU,
size = problemsToSolve[problemsReadingIndex].size,
isSynchronizable = gpuResultsIsSynchronizable[stream][i],
shortestSynchronizingWordLength = gpuResultsShortestSynchronizingWordLength[stream][i],
algorithmName = GetType().Name
}
}
;
if (gpuResultsIsSynchronizable[stream][i] && gpuResultsShortestSynchronizingWordLength[stream][i] > maximumPermissibleWordLength)
{
CernyConjectureFailingIndex = resultsWritingIndex + offset + i;
break;
}
}
});
foreach (var arrays in new IEnumerable <Array>[] { gpuAs, gpuBs, isSynchronizable, shortestSynchronizingWordLength })
{
foreach (var array in arrays)
{
Gpu.Free(array);
}
}
Gpu.FreeAllImplicitMemory();
foreach (var stream in streams)
{
stream.Dispose();
}
#if (benchmark)
computationResults[resultsWritingIndex].benchmarkResult = new BenchmarkResult
{
benchmarkedTime = benchmarkTiming.Elapsed,
totalTime = totalTiming.Elapsed
};
#endif
return(CernyConjectureFailingIndex);
}
public ComputationResult[] Compute(
Problem[] problemsToSolve,
int streamCount,
Action asyncAction = null,
int warpCount = 16)
// cannot be more warps since more memory should be allocated
{
#if (benchmark)
var totalTiming = new Stopwatch();
totalTiming.Start();
var benchmarkTiming = new Stopwatch();
#endif
var gpu = Gpu.Default;
var n = problemsToSolve.First().size;
var power = 1 << n;
var maximumPermissibleWordLength = (n - 1) * (n - 1);
var maximumWarps = gpu.Device.Attributes.MaxThreadsPerBlock / gpu.Device.Attributes.WarpSize;
if (warpCount > maximumWarps)
{
warpCount = maximumWarps;
}
var problemsPerStream = (problemsToSolve.Count() + streamCount - 1) / streamCount;
var problemsPartitioned = Enumerable.Range(0, streamCount)
.Select(i => problemsToSolve.Skip(problemsPerStream * i)
.Take(problemsPerStream)
.ToArray())
.Where(partition => partition.Length > 0)
.ToArray();
streamCount = problemsPartitioned.Length;
var streams = Enumerable.Range(0, streamCount)
.Select(_ => gpu.CreateStream()).ToArray();
var gpuA = problemsPartitioned.Select(problems => gpu.Allocate <int>(problems.Length * n)).ToArray();
var gpuB = problemsPartitioned.Select(problems => gpu.Allocate <int>(problems.Length * n)).ToArray();
var isSynchronizable = problemsPartitioned.Select(problems => gpu.Allocate <bool>(problems.Length)).ToArray();
gpu.Copy(n, problemSize);
var launchParameters = new LaunchParam(
new dim3(1, 1, 1),
new dim3(gpu.Device.Attributes.WarpSize * warpCount, 1, 1),
2 * n * sizeof(ushort)
);
var gpuResultsIsSynchronizable = problemsPartitioned
.Select(problems => new bool[problems.Length])
.ToArray();
for (int stream = 0; stream < streamCount; stream++)
{
var problems = problemsPartitioned[stream];
var matrixA = new int[problems.Length * n];
var matrixB = new int[problems.Length * n];
Parallel.For(0, problems.Length, problem =>
{
Array.ConstrainedCopy(problems[problem].stateTransitioningMatrixA, 0, matrixA, problem * n, n);
Array.ConstrainedCopy(problems[problem].stateTransitioningMatrixB, 0, matrixB, problem * n, n);
});
streams[stream].Copy(matrixA, gpuA[stream]);
streams[stream].Copy(matrixB, gpuB[stream]);
// TODO: change this entirely
// warning, this might not compute the length of a shortest synchronizing word but it will verify the Cerny conjecture
streams[stream].Launch(
Kernel,
launchParameters,
gpuA[stream],
gpuB[stream],
isSynchronizable[stream]
);
}
asyncAction?.Invoke();
for (int stream = 0; stream < streamCount; stream++)
{
#if (benchmark)
benchmarkTiming.Start();
#endif
streams[stream].Synchronize();
#if (benchmark)
benchmarkTiming.Stop();
#endif
streams[stream].Copy(isSynchronizable[stream], gpuResultsIsSynchronizable[stream]);
}
gpu.Synchronize();
#if (benchmark)
#endif
//Enumerable.Range(0, streamCount)
// .SelectMany(stream => gpuResultsIsSynchronizable[stream])
// .Select((result) =>
// {
// });
var results = new ComputationResult[problemsToSolve.Count()];
var slimCPU = new SlimCPU();
var listOfCPUProblems = new List <Problem>();
var cpuProblemIndex = new List <int>();
int generalIndex = 0;
for (int stream = 0; stream < streamCount; stream++)
{
for (int index = 0; index < gpuResultsIsSynchronizable[stream].Length; index++)
{
if (gpuResultsIsSynchronizable[stream][index])
{
results[generalIndex] = new ComputationResult
{
isSynchronizable = true,
computationType = ComputationType.CPU_GPU_Combined,
size = n,
algorithmName = GetType().Name
};
}
else
{
listOfCPUProblems.Add(problemsPartitioned[stream][index]);
cpuProblemIndex.Add(generalIndex);
}
generalIndex++;
}
}
var cpuResults = slimCPU.Compute(listOfCPUProblems.ToArray(), slimCPU.GetBestParallelism());
for (int i = 0; i < listOfCPUProblems.Count; i++)
{
results[cpuProblemIndex[i]] = cpuResults[i];
}
if (cpuResults.Any(result => result.isSynchronizable && result.shortestSynchronizingWordLength > maximumPermissibleWordLength))
{
throw new Exception("Cerny conjecture is false");
}
foreach (var array in gpuA.AsEnumerable <Array>()
.Concat(gpuB)
.Concat(isSynchronizable))
{
Gpu.Free(array);
}
foreach (var stream in streams)
{
stream.Dispose();
}
#if (benchmark)
results[0].benchmarkResult = new BenchmarkResult
{
benchmarkedTime = benchmarkTiming.Elapsed,
totalTime = totalTiming.Elapsed
};
#endif
return(results);
}
protected virtual void FreeCache()
{
Free();
Gpu.Free(rneuronsCacheArr);
}
