/// <summary>Returns the matrix product M1*M2</summary>
/// <param name="M1">First matrix</param>
/// <param name="M2">Second matrix</param>
public float[,] MultiplyLocals(float[,] M1, float[,] M2)
{
//M pxq, N qxr
int p = M1.GetLength(0);
int q = M1.GetLength(1);
int r = M2.GetLength(1);
if (q != M2.GetLength(0)) throw new Exception("Matrix dimensions do not match for multiplication");
float[] vecM1 = MatrixToVector(M1, ref p, ref q);
float[] vecM2 = MatrixToVector(M2, ref q, ref r);
float[] vecResp = new float[p * r];
CLCalc.Program.Variable varResp = new CLCalc.Program.Variable(vecResp);
CLCalc.Program.Variable varM1 = new CLCalc.Program.Variable(vecM1);
CLCalc.Program.Variable varM2 = new CLCalc.Program.Variable(vecM2);
//Finaliza a soma dos elementos
int[] vecQ = new int[1] { q };
CLCalc.Program.Variable varQ = new CLCalc.Program.Variable(vecQ);
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[4] { varResp, varM1, varM2, varQ };
int[] max = new int[2] { p, r };
floatMatrixMultLocals.Execute(args, max, new int[] { 8, 8 });
varResp.ReadFromDeviceTo(vecResp);
varResp.Dispose();
return VectorToMatrix(vecResp, ref p, ref r);
}
static void Main()
{
//Initializes OpenCL Platforms and Devices and sets everything up
CLCalc.InitCL();
//Create vectors with 2000 numbers
float[] v1 = new float[n], v2 = new float[n];
var vResult = new float[n];
//Creates population for v1 and v2
for (int i = 0; i < n; i++)
{
v1[i] = (float)i / 10;
v2[i] = -(float)i / 9;
}
//var prog = new ComputeProgram(CLCalc.Program.Context, "");
//Compiles the source codes. The source is a string array because the user may want
//to split the source into many strings.
CLCalc.Program.Compile(new string[] { vecSum });
//Gets host access to the OpenCL floatVectorSum kernel
CLCalc.Program.Kernel VectorSum = new CLCalc.Program.Kernel("floatVectorSum");
//Creates vectors v1 and v2 in the device memory
CLCalc.Program.Variable varV1 = new CLCalc.Program.Variable(v1);
CLCalc.Program.Variable varV2 = new CLCalc.Program.Variable(v2);
//Arguments of VectorSum kernel
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { varV1, varV2 };
//How many workers will there be? We need "n", one for each element
int[] workers = new int[1] { n };
sw.Start();
//Execute the kernel
for (int i = 0; i < count; i++) DoOCL(VectorSum, args, workers);
sw.Stop();
//Read device memory varV1 to host memory vResult
varV1.ReadFromDeviceTo(vResult);
Console.WriteLine("OpenCL: {0}", sw.ElapsedTicks);
sw.Restart();
for (int i = 0; i < count; i++) DoCPU(v1, v2, vResult);
sw.Stop();
Console.WriteLine("CPU: {0}", sw.ElapsedTicks);
PressAny();
}
/// <summary>ImageData constructor. Reads data from a bitmap</summary>
/// <param name="bmp">Bitmap to read from</param>
public ImageData(Bitmap bmp)
{
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown) CLCalc.InitCL();
width = bmp.Width;
height = bmp.Height;
//Allocates space for data
Data = new byte[3 * width * height];
//Reads bmp to local Data variable
ReadToLocalData(bmp);
//Transfer data to OpenCL device
varData = new CLCalc.Program.Variable(Data);
}
/// <summary>Computes the Discrete Fourier Transform of a double2 vector x whose length is a power of 16.
/// x = { Re[x0] Im[x0] Re[x1] Im[x1] ... Re[xn] Im[xn] }, n = power of 16 (Length = 2*pow(16,n))</summary>
public static double[] FFT16(double[] x)
{
if (CLx == null || CLx.OriginalVarLength != x.Length)
{
CLx = new CLCalc.Program.Variable(x);
CLy = new CLCalc.Program.Variable(x);
}
//Writes original content
CLx.WriteToDevice(x);
CLy = FFT16(ref CLx);
double[] y = new double[x.Length];
CLy.ReadFromDeviceTo(y);
return y;
}
/// <summary>Computes the Discrete Fourier Transform of a double2 vector x whose length is a power of 16.
/// x = { Re[x0] Im[x0] Re[x1] Im[x1] ... Re[xn] Im[xn] }, n = power of 16 (Length = 2*pow(16,n))</summary>
public static CLCalc.Program.Variable FFT16(ref CLCalc.Program.Variable CLx)
{
if (CLy == null || CLy.OriginalVarLength != CLx.OriginalVarLength) CLy = new CLCalc.Program.Variable(new float[CLx.OriginalVarLength]);
if (CLCalc.CLAcceleration != CLCalc.CLAccelerationType.UsingCL) return null;
int nn = (int)Math.Log(CLx.OriginalVarLength >> 1, 16);
nn = 1 << ((nn << 2) + 1);
if (nn != CLx.OriginalVarLength) throw new Exception("Number of elements should be a power of 16 ( vector length should be 2*pow(16,n) )");
if (kernelfft_radix16 == null)
{
InitKernels();
}
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { CLx, CLy, CLp };
CLCalc.Program.Variable[] args2 = new CLCalc.Program.Variable[] { CLy, CLx, CLp };
bool usar2 = true;
int[] p = new int[] { 1 };
CLp.WriteToDevice(p);
int n = CLx.OriginalVarLength >> 5;
while (p[0] <= n)
{
usar2 = !usar2;
if (usar2)
kernelfft_radix16.Execute(args2, n);
else
kernelfft_radix16.Execute(args, n);
p[0] = p[0] << 4;
CLp.WriteToDevice(p);
}
if (usar2)
{
CLCalc.Program.Variable temp = CLx;
CLx = CLy; CLy = temp;
}
return CLy;
}
/// <summary>Applies given filter to the image</summary>
/// <param name="imgDt">Image to be filtered</param>
/// <param name="Filter">Filter. [3*size*size]</param>
public static void ApplyFilter(ImageData imgDt, float[] Filter, bool useOpenCL, bool useWorkDim2)
{
int FilterSize = (int)Math.Sqrt(Filter.Length/3);
if (Filter.Length != 3 * FilterSize * FilterSize)
throw new Exception("Invalid filter");
if (!Initialized && useOpenCL) Init(FilterSize);
//Writes filter to device
if(useOpenCL) varFilter.WriteToDevice(Filter);
if (FilteredVals == null || FilteredVals.Length != imgDt.Height * imgDt.Width * 3)
{
//Filtered values
FilteredVals = new float[imgDt.Height * imgDt.Width * 3];
varFiltered = new CLCalc.Program.Variable(FilteredVals);
}
//Width
if (useOpenCL) varWidth.WriteToDevice(new int[] { imgDt.Width });
//Executes filtering
int mean = (FilterSize - 1) / 2;
if (useOpenCL)
{
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { imgDt.varData, varFilter, varFiltered, varWidth };
if (useWorkDim2)
{
kernelApplyFilterWorkDim2.Execute(args, new int[] { imgDt.Width - FilterSize, imgDt.Height - FilterSize });
}
else
{
kernelApplyFilter.Execute(args, new int[] { imgDt.Height - FilterSize });
}
//Reads data back
varFiltered.ReadFromDeviceTo(FilteredVals);
}
else
{
ApplyFilter(imgDt.Data, Filter, FilteredVals, new int[] { imgDt.Width }, imgDt.Height - FilterSize);
}
//Writes to image data
for (int y = mean; y < imgDt.Height - mean - 1; y++)
{
int wy = imgDt.Width * y;
for (int x = mean; x < imgDt.Width - mean - 1; x++)
{
int ind = 3 * (x + wy);
imgDt.Data[ind] = (byte)FilteredVals[ind];
imgDt.Data[ind + 1] = (byte)FilteredVals[ind + 1];
imgDt.Data[ind + 2] = (byte)FilteredVals[ind + 2];
}
}
//Writes filtered values
//In the future this rewriting can be avoided
//because byte_addressable will be widely available
if (useOpenCL) imgDt.varData.WriteToDevice(imgDt.Data);
}
/// <summary>Initializes class</summary>
private static void Init(int FilterSize)
{
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown) CLCalc.InitCL();
//Compiles source code
CLCalc.Program.Compile((new CLFilterSrc()).src);
//Creates kernel
kernelApplyFilter = new CLCalc.Program.Kernel("ApplyFilter");
kernelApplyFilterWorkDim2 = new CLCalc.Program.Kernel("ImgFilter");
//Creates filter
varFilter = new CLCalc.Program.Variable(new float[3 * FilterSize * FilterSize]);
//Width
varWidth = new CLCalc.Program.Variable(new int[1]);
Initialized = true;
}
/// <summary>New matrix constructor</summary>
/// <param name="Vals">Matrix values</param>
public floatMatrix(float[,] Vals)
{
nRows = Vals.GetLength(0);
nCols = Vals.GetLength(1);
Values = new float[nRows * nCols];
if (UseOpenCLIfAvailable && CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
CLValues = new CLCalc.Program.Variable(Values);
CLDim = new CLCalc.Program.Variable(new int[] { nRows, nCols });
CLCoef = new CLCalc.Program.Variable(new float[1]);
}
SetValues(Vals);
}
/// <summary>Constructor.</summary>
public SparseLinalg()
{
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown)
{
try { CLCalc.InitCL(); }
catch
{
}
}
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
//Creates control variables
dprod = new float[SparseLinalg.GLOBALWORKSIZE];
dotProd = new CLCalc.Program.Variable(dprod);
dotProdSum = new CLCalc.Program.Variable(new float[1]);
int[] i = new int[1];
vLenBy4 = new CLCalc.Program.Variable(i);
CLNonZeroElemsPerRow = new CLCalc.Program.Variable(new int[1]);
}
}
private static void InitKernels()
{
string s = new CLFFTSrc().s;
CLCalc.InitCL();
try
{
CLCalc.Program.Compile(s);
}
catch
{
}
kernelfft_radix16 = new CLCalc.Program.Kernel("fft_radix16");
kernelfft_radix4 = new CLCalc.Program.Kernel("fft_radix4");
kernelConjugate = new CLCalc.Program.Kernel("Conjugate");
CLp = new CLCalc.Program.Variable(new int[1]);
}
/// <summary>Computes the inverse Discrete Fourier Transform of a float2 vector x whose length is a power of 4.
/// x = { Re[x0] Im[x0] Re[x1] Im[x1] ... Re[xn] Im[xn] }, n = power of 4 (Length = 2*pow(4,n))</summary>
public static float[] iFFT4(float[] x)
{
if (CLx == null || CLx.OriginalVarLength != x.Length)
{
CLx = new CLCalc.Program.Variable(x);
CLy = new CLCalc.Program.Variable(x);
}
//Writes original content
CLx.WriteToDevice(x);
CLy = iFFT4(CLx);
float[] y = new float[x.Length];
CLy.ReadFromDeviceTo(y);
return y;
}
/// <summary>Constructor.</summary>
/// <param name="InitialState">Initial state of system</param>
/// <param name="StepSize">Desired step per integration pass</param>
/// <param name="InitialIndepVarValue">Initial independent variable value</param>
/// <param name="DerivativeCalculator">Function to calculate derivatives vector</param>
public doubleODE46(double InitialIndepVarValue, double StepSize, double[] InitialState, DerivCalcDeleg DerivativeCalculator)
{
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown)
{
CLCalc.InitCL();
}
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.NotUsingCL)
throw new Exception("OpenCL not available");
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
ODE46Source Source = new ODE46Source();
string[] s = new string[] { @"
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
", Source.doubleStep2, Source.doubleStep3, Source.doubleStep4, Source.doubleStep5, Source.doubleStep6, Source.doubleFinalizeCalc };
CLCalc.Program.Compile(s);
//Calculador de derivada
Derivs = DerivativeCalculator;
//Scalars
double[] xx = new double[1] { InitialIndepVarValue };
x = new CLCalc.Program.Variable(xx);
xsav = new CLCalc.Program.Variable(xx);
//Sets initial values to Device and local variables
hdid = new CLCalc.Program.Variable(xx);
currentX = InitialIndepVarValue;
SetStep(StepSize);
//Vectors
yy = new double[InitialState.Length];
for (int i = 0; i < InitialState.Length; i++) yy[i] = InitialState[i];
ysav = new CLCalc.Program.Variable(yy);
k1 = new CLCalc.Program.Variable(InitialState);
k2 = new CLCalc.Program.Variable(InitialState);
k3 = new CLCalc.Program.Variable(InitialState);
k4 = new CLCalc.Program.Variable(InitialState);
k5 = new CLCalc.Program.Variable(InitialState);
k6 = new CLCalc.Program.Variable(InitialState);
absError = new CLCalc.Program.Variable(new double[InitialState.Length]);
y = new CLCalc.Program.Variable(yy);
//Kernels
KernelFinalizeCalc = new CLCalc.Program.Kernel("doubleFinalizeCalc");
KernelUpdateX = new CLCalc.Program.Kernel("doubleUpdateX");
KernelRK46YStep2 = new CLCalc.Program.Kernel("doubleYStep2");
KernelRK46XStep2 = new CLCalc.Program.Kernel("doubleXStep2");
KernelRK46YStep3 = new CLCalc.Program.Kernel("doubleYStep3");
KernelRK46XStep3 = new CLCalc.Program.Kernel("doubleXStep3");
KernelRK46YStep4 = new CLCalc.Program.Kernel("doubleYStep4");
KernelRK46XStep4 = new CLCalc.Program.Kernel("doubleXStep4");
KernelRK46YStep5 = new CLCalc.Program.Kernel("doubleYStep5");
KernelRK46XStep5 = new CLCalc.Program.Kernel("doubleXStep5");
KernelRK46YStep6 = new CLCalc.Program.Kernel("doubleYStep6");
KernelRK46XStep6 = new CLCalc.Program.Kernel("doubleXStep6");
//Kernel arguments
ArgsFinalize = new CLCalc.Program.Variable[] { x, hdid, y, ysav, absError, k1, k2, k3, k4, k5, k6 };
ArgsRK46Y = new CLCalc.Program.Variable[] { x, hdid, y, ysav, k1, k2, k3, k4, k5, k6 };
ArgsRK46X = new CLCalc.Program.Variable[] { x, hdid, xsav };
NStates = new int[1] { InitialState.Length };
NScalar = new int[1] { 1 };
//Data retrieving
yerr = new double[NStates[0]];
xRet = new double[NScalar[0]];
}
}
/// <summary>Creates vector from M elements sequentially</summary>
/// <param name="symM">Symmetric matrix to use</param>
public floatVector(floatSymPosDefMatrix symM)
{
this.CLValues = symM.CLValues;
this.Values = symM.Values;
//Since I'm probably going to modify the matrix, I want a new Cholesky factorization
//if I ever call a LinearSolve
symM.IsCholeskyFactorized = false;
if (UseOpenCLIfAvailable && CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
CLCoef = new CLCalc.Program.Variable(new float[1]);
}
}
/// <summary>OpenCL vector constructor</summary>
/// <param name="Vals">Vector elements</param>
public floatVector(float[] Vals)
{
this.Values = (float[])Vals.Clone();
if (UseOpenCLIfAvailable && CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
CLValues = new CLCalc.Program.Variable(Values);
CLCoef = new CLCalc.Program.Variable(new float[1]);
}
}
private void LocalInitCL()
{
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown) CLCalc.InitCL();
if (UseOpenCLIfAvailable && CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
CLoffSet = new CLCalc.Program.Variable(new int[1]);
CLValues = new CLCalc.Program.Variable(this.Values);
invL11 = new float[(SUBMATRIXSIZE * (SUBMATRIXSIZE + 1)) >> 1];
CLinvl11 = new CLCalc.Program.Variable(invL11);
int NMultiple = N;
if (N % SUBMATRIXSIZE != 0)
{
NMultiple = N / SUBMATRIXSIZE;
NMultiple = SUBMATRIXSIZE * (NMultiple + 1);
cholDec = new float[(NMultiple * (NMultiple + 1)) >> 1];
for (int i = 0; i < Values.Length; i++) cholDec[i] = Values[i];
}
else
{
cholDec = (float[])this.Values.Clone();
}
CLcholDec = new CLCalc.Program.Variable(cholDec);
CLprevVals = new CLCalc.Program.Variable(new float[N]);
CLb = new CLCalc.Program.Variable(new float[N]);
CLy = new CLCalc.Program.Variable(new float[N]);
CLn = new CLCalc.Program.Variable(new int[] { N });
}
}
/// <summary>Backsubstitutes to solve a linear system with a matrix right hand size</summary>
private void LinsolveCLMatrix(floatMatrix M, ref floatMatrix resp)
{
//System.Diagnostics.Stopwatch sw = new System.Diagnostics.Stopwatch();
//System.Diagnostics.Stopwatch sw1 = new System.Diagnostics.Stopwatch();
//sw.Start();
//number of RHS as multiple of SUBMATRIXSIZE
int nRHSMult = M.Rows / SUBMATRIXSIZE;
int nRHSleftOver = M.Rows - SUBMATRIXSIZE*nRHSMult;
if (!UseOpenCLIfAvailable || CLCalc.CLAcceleration != CLCalc.CLAccelerationType.UsingCL)
{
linsolveMatrix(M, ref resp);
return;
}
//Copy elements to CLb
if (CLb == null || CLb.OriginalVarLength < M.Values.Length)
{
CLb = new CLCalc.Program.Variable(M.Values);
CLy = new CLCalc.Program.Variable(M.Values);
}
kernelCopyBuffer.Execute(new CLCalc.Program.MemoryObject[] { M.CLValues, CLb }, M.Values.Length);
int nEqs = M.Rows;
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { CLcholDec, CLy, CLb, CLoffSet, CLn };
int[] offset = new int[1];
//DEBUG
//float[] yDebug = new float[M.Values.Length];
//float[] bDebug = new float[M.Values.Length];
//this.CLcholDec.ReadFromDeviceTo(cholDec);
//Forward substitution
int i;
for (i = 0; i < N; i += SUBMATRIXSIZE)
{
offset[0] = i;
CLoffSet.WriteToDevice(offset);
int size = Math.Min(SUBMATRIXSIZE, N - i);
kernelFwdUpperBackSubs.Execute(args, new int[] { size, nEqs }, new int[] { size, 1 });
////DEBUG
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
//sw1.Start();
//propagation
if (i + SUBMATRIXSIZE < N)
{
if (nRHSMult > 0) kernelFwdPropag.Execute(args, new int[] { N - i - SUBMATRIXSIZE, nRHSMult * SUBMATRIXSIZE }, new int[] { 1, SUBMATRIXSIZE });
if (nRHSleftOver > 0)
kernelFwdPropag2.Execute(args, new int[] { N - i - SUBMATRIXSIZE, nRHSleftOver }, new int[] { 1, nRHSleftOver }, new int[] { 0, nRHSMult * SUBMATRIXSIZE });
}
//OpenCLTemplate.CLCalc.Program.CommQueues[OpenCLTemplate.CLCalc.Program.DefaultCQ].Finish();
//sw1.Stop();
////DEBUG
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
}
//Backward subst. Stores answer in CLb
args = new CLCalc.Program.Variable[] { CLcholDec, CLb, CLy, CLoffSet, CLn };
//Backward substitution
for (i = N - SUBMATRIXSIZE; i >= 0; i -= SUBMATRIXSIZE)
{
offset[0] = i;
CLoffSet.WriteToDevice(offset);
int size = SUBMATRIXSIZE;
kernelBkLowerBackSubs.Execute(args, new int[] { size, nEqs }, new int[] { size, 1 });
if (i > 0)
{
//Propagation using __local storage
if (nRHSMult > 0) kernelBackPropag.Execute(args, new int[] { i, nRHSMult * SUBMATRIXSIZE }, new int[] { 1, SUBMATRIXSIZE });
//leftovers (not multiples of SUBMATRIXSIZE)
if (nRHSleftOver > 0)
kernelBackPropag2.Execute(args, new int[] { i, nRHSleftOver }, new int[] { 1, nRHSleftOver }, new int[] { 0, nRHSMult * SUBMATRIXSIZE });
}
}
if (SUBMATRIXSIZE + i > 0)
{
offset[0] = 0; CLoffSet.WriteToDevice(offset);
kernelBkLowerBackSubs.Execute(args, new int[] { SUBMATRIXSIZE + i, nEqs }, new int[] { SUBMATRIXSIZE + i, 1 });
}
kernelCopyBuffer.Execute(new CLCalc.Program.Variable[] { CLb, resp.CLValues }, resp.Values.Length);
//OpenCLTemplate.CLCalc.Program.CommQueues[OpenCLTemplate.CLCalc.Program.DefaultCQ].Finish();
//sw.Stop();
}
private void linsolveCL(floatVector CLbb, ref floatVector resp)
{
if (!UseOpenCLIfAvailable || CLCalc.CLAcceleration != CLCalc.CLAccelerationType.UsingCL)
{
linsolve(CLbb.Values, ref resp);
return;
}
//int NMultiple = N;
////float[] bAugm;
//if (N % SUBMATRIXSIZE != 0)
//{
// NMultiple = N / SUBMATRIXSIZE;
// NMultiple = SUBMATRIXSIZE * (NMultiple + 1);
//}
////bAugm = new float[NMultiple];
////for (int i = 0; i < bb.Length; i++) bAugm[i] = bb[i];
if (resp == null) resp = new floatVector(new float[N]);
//Copy elements to CLb
if (CLb == null || CLb.OriginalVarLength < CLbb.Length) CLb = new CLCalc.Program.Variable(CLbb.Values);
kernelCopyBuffer.Execute(new CLCalc.Program.MemoryObject[] { CLbb.CLValues, CLb }, CLbb.Length);
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { CLcholDec, CLy, CLb, CLoffSet, CLn };
int[] offset = new int[1];
//float[] yDebug = new float[N];
//float[] bDebug = new float[N];
//Forward substitution
int i;
for (i = 0; i < N; i += SUBMATRIXSIZE)
{
offset[0] = i;
CLoffSet.WriteToDevice(offset);
int size = Math.Min(SUBMATRIXSIZE, N - i);
kernelFwdUpperBackSubs.Execute(args, new int[] { size }, new int[] { size });
////DEBUG
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
//propagation
if (i + SUBMATRIXSIZE < N)
{
kernelFwdPropag.Execute(args, N - i - SUBMATRIXSIZE);
}
////DEBUG
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
//CLcholDec.ReadFromDeviceTo(cholDec);
}
//Backward subst. Stores answer in CLb
args = new CLCalc.Program.Variable[] { CLcholDec, CLb, CLy, CLoffSet, CLn };
//Backward substitution
for (i = N - SUBMATRIXSIZE; i >= 0; i -= SUBMATRIXSIZE)
{
offset[0] = i;
CLoffSet.WriteToDevice(offset);
int size = SUBMATRIXSIZE;
kernelBkLowerBackSubs.Execute(args, new int[] { size }, new int[] { size });
////DEBUG
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
if (i > 0)
{
kernelBackPropag.Execute(args, i);
}
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
}
if (SUBMATRIXSIZE + i > 0)
{
offset[0] = 0; CLoffSet.WriteToDevice(offset);
kernelBkLowerBackSubs.Execute(args, new int[] { SUBMATRIXSIZE + i }, new int[] { SUBMATRIXSIZE + i });
}
//CLy.ReadFromDeviceTo(yDebug);
//CLb.ReadFromDeviceTo(bDebug);
kernelCopyBuffer.Execute(new CLCalc.Program.Variable[] { CLb, resp.CLValues }, N);
}
/// <summary>Cholesky decomposition using OpenCL with Blocks</summary>
public void CLBlockCholesky()
{
//If matrix dimension is not a multiple of SUBMATRIXSIZE
//pad with zeros.
int NMultiple = N;
if (N % SUBMATRIXSIZE != 0)
{
NMultiple = N / SUBMATRIXSIZE;
NMultiple = SUBMATRIXSIZE * (NMultiple + 1);
}
if (!IsMatrixInClMemoryUpdated)
{
for (int i = 0; i < Values.Length; i++) cholDec[i] = Values[i];
CLcholDec.WriteToDevice(cholDec);
}
else
{
kernelCopyBuffer.Execute(new CLCalc.Program.MemoryObject[] { CLValues, CLcholDec }, CLValues.OriginalVarLength);
}
int SubMatrixSize = SUBMATRIXSIZE;
int GlobalSize;
//Important. Set offset to zero
CLoffSet.WriteToDevice(new int[] { 0 });
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { CLcholDec, CLoffSet, CLinvl11 };
GlobalSize = (SubMatrixSize * (SubMatrixSize + 1)) >> 1;
for (int i = 0; i < NMultiple; i += SubMatrixSize)
{
//Computes Cholesky factor L11 and its inverse
kernelCholeskyDiagBlock.Execute(args, new int[] { GlobalSize }, new int[] { GlobalSize });
//CLcholDec.ReadFromDeviceTo(cholDec);
//Computes column panel L21
//Note: offSet has been updated, kernel should use its value-1
//Number of submatrices to update: (N-i)/SubMatrixSize
int nSubMatrices = (NMultiple - i) / SubMatrixSize - 1;
if (nSubMatrices > 0)
{
//Computes panels and updates main diagonals
kernelCholeskyComputePanel.Execute(args, new int[] { nSubMatrices * SubMatrixSize, SubMatrixSize }, new int[] { SubMatrixSize, SubMatrixSize });
//CLcholDec.ReadFromDeviceTo(cholDec);
//Still need to update nSubMatrices*(nSubMatrices-1)/2 full matrices in the Cholesky decomposition
//They start at indexes [i+SubMatrixSize i], and they are the offdiagonal block matrices
int totalSubMatricesToUpdate = ((nSubMatrices - 1) * nSubMatrices) >> 1;
if (totalSubMatricesToUpdate > 0)
{
kernelCholeskyForwardProp.Execute(args, new int[] { totalSubMatricesToUpdate * SubMatrixSize, SubMatrixSize }, new int[] { SubMatrixSize, SubMatrixSize });
}
}
//CLcholDec.ReadFromDeviceTo(cholDec);
}
//CLcholDec.ReadFromDeviceTo(cholDec);
this.IsCholeskyFactorized = true;
}
/// <summary>Solves linear system Mx = b using conjugate gradient method. Doesn't try to improve the solution obtained.</summary>
/// <param name="M">Matrix M</param>
/// <param name="b">Vector b</param>
/// <param name="tol">Error tolerance</param>
/// <param name="x">Initial guess</param>
public void LinSolveCLStep(CLImgSparseMatrix M, CLImgVector b, float tol, ref CLImgVector x)
{
int n = b.Length;
int nBy4 = 1 + ((n - 1) >> 2);
if (lambda == null)
{
lambda = new float[1];
CLlambda = new CLCalc.Program.Variable(lambda);
}
if (r == null || r.Length != n)
{
r = new CLImgVector(n);
p = new CLImgVector(n);
//x = new CLImgVector(n);
Ap = new CLImgVector(n);
temp = new CLImgVector(n);
}
if (temp == null) temp = new CLImgVector(n);
if (x == null || x.Length != n) x = new CLImgVector(n);
float alpha, beta, RDotROld, RDotR;
//Initialization
Multiply(M, x, Ap);
CLCalc.Program.MemoryObject[] args = new CLCalc.Program.MemoryObject[] { b.CLVector, Ap.CLVector, r.CLVector, p.CLVector };
kernelInitRP.Execute(args, nBy4);
//Loop
int count = 0;
RDotR = DotProduct(r, r);
while (count<1 || ((RDotR > tol) && (count < n*MAXITER)))
{
RDotROld = RDotR;
//if ((count & 0x0080) == 0)
//{
// Multiply(M, x, Ap);
// args = new CLCalc.Program.MemoryObject[] { b.CLVector, Ap.CLVector, r.CLVector, p.CLVector };
// kernelInitRP.Execute(args, nBy4);
//}
Multiply(M, p, Ap);
alpha = RDotROld / DotProduct(Ap, p);
//Update x
kernelCopyToTemp.Execute(new CLCalc.Program.MemoryObject[] { x.CLVector, temp.CLVector }, nBy4);
lambda[0] = alpha; CLlambda.WriteToDevice(lambda);
kernelMultiplyAdd.Execute(new CLCalc.Program.MemoryObject[] { CLlambda, p.CLVector, temp.CLVector, x.CLVector }, nBy4);
//Update r
kernelCopyToTemp.Execute(new CLCalc.Program.MemoryObject[] { r.CLVector, temp.CLVector }, nBy4);
lambda[0] = -alpha; CLlambda.WriteToDevice(lambda);
kernelMultiplyAdd.Execute(new CLCalc.Program.MemoryObject[] { CLlambda, Ap.CLVector, temp.CLVector, r.CLVector }, nBy4);
RDotR = DotProduct(r, r);
beta = RDotR / RDotROld;
//Update p
kernelCopyToTemp.Execute(new CLCalc.Program.MemoryObject[] { p.CLVector, temp.CLVector }, nBy4);
lambda[0] = beta; CLlambda.WriteToDevice(lambda);
kernelMultiplyAdd.Execute(new CLCalc.Program.MemoryObject[] { CLlambda, temp.CLVector, r.CLVector, p.CLVector }, nBy4);
count++;
}
}
/// <summary>Calculates LU decomposition of M matrix</summary>
/// <param name="M">Matrix to decompose</param>
/// <param name="n">Matrix dimension</param>
/// <param name="varindx">Swap index</param>
private CLCalc.Program.Variable LUDecomp(double[,] M, int n, out CLCalc.Program.Variable varindx)
{
//arguments and work_dim
CLCalc.Program.Variable[] args;
int[] max;
//Matrix to vector
double[] vecM = MatrixToVector(M, ref n, ref n);
CLCalc.Program.Variable varM = new Program.Variable(vecM);
//Scaling transformation
double[] vv = new double[n];
CLCalc.Program.Variable varvv = new Program.Variable(vv);
max = new int[1] { n };
args = new CLCalc.Program.Variable[] { varM, varvv };
doubleLUScale.Execute(args, max);
//In order LU factorization (Crout)
int[] J = new int[1] { 0 };
CLCalc.Program.Variable varJ = new Program.Variable(J);
int[] N = new int[1] { n };
CLCalc.Program.Variable varN = new Program.Variable(N);
int[] indx = new int[n];
varindx = new Program.Variable(indx);
args = new Program.Variable[] { varM, varJ, varN, varindx, varvv };
for (J[0] = 0; J[0] < n; J[0]++)
{
varJ.WriteToDevice(J);
max[0] = J[0];
doubleLUCalcBetas.Execute(args, max);
max[0] = n - J[0];
doubleLUCalcAlphas.Execute(args, max);
max[0] = 1;
doubleLUCalcPivo.Execute(args, max);
max[0] = n;
doubleLUTrocaCols.Execute(args, max);
if (J[0] != n - 1)
{
max[0] = n - J[0] - 1;
doubleLUDivByPivot.Execute(args, max);
}
}
return varM;
}
/// <summary>Constructor.</summary>
/// <param name="nMasses">Number of masses in the system</param>
/// <param name="nConnections">Number of connections</param>
/// <param name="Masses">Mass of each vertex</param>
/// <param name="InitialStateSpace">Position and velocity of vertexes
/// [2*3*i] - posx, [2*(3*i+1)] - posy, [2*(3*i+2)] - posz,
/// [1+2*3*i] - velx, [1+2*(3*i+1)] - vely, [1+2*(3*i+2)] - velz</param>
/// <param name="Origins">Origin vertex of connections. Spring connects Origin[i] to Dests[i]</param>
/// <param name="Dests">Destination vertex of connections. Spring connects Origin[i] to Dests[i]</param>
/// <param name="SpringKs">Spring constant for each connection</param>
/// <param name="GroundKs">Spring constant for each mass, connecting to ground (nMass)</param>
/// <param name="Damp">Structural damping (relative-speed dependant) (nConnections)</param>
/// <param name="GroundDamp">Absolute damping proportional to speed relative to Earth (nMass)</param>
public floatDEM(int nMasses, int nConnections,
float[] Masses, float[] InitialStateSpace,
int[] Origins, int[] Dests, float[] SpringKs, float[] Damp, float[] GroundKs, float[] GroundDamp)
{
#region Consistency check
if (Masses.Length != nMasses)
throw new Exception("Invalid Masses length (!=nMasses)");
if (InitialStateSpace.Length != 6 * nMasses)
throw new Exception("Invalid positions length (!=6*nMasses - x, y, z)");
if (Origins.Length != nConnections)
throw new Exception("Invalid Origins length (!=nConnections)");
if (Dests.Length != nConnections)
throw new Exception("Invalid Dests length (!=nConnections)");
if (SpringKs.Length != nConnections)
throw new Exception("Invalid SpringKs length (!=nConnections)");
if (GroundKs.Length != nMasses)
throw new Exception("Invalid GroundKs length (!=nMasses)");
if (Damp.Length != nConnections)
throw new Exception("Invalid Damp length (!=nConnections)");
if (GroundDamp.Length != nMasses)
throw new Exception("Invalid GroundDamp length (!=nMasses)");
#endregion
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown)
{
CLCalc.InitCL();
}
#region Variables reading
//Sizes
nConn = new int[1] { nConnections };
nM = new int[1] { nMasses };
//Inputs
m = new Program.Variable(Masses);
float[] InitialPositions = new float[3 * nMasses];
for (int i = 0; i < 3 * nMasses; i++) InitialPositions[i] = InitialStateSpace[2 * i];
posOrig = new Program.Variable(InitialPositions);
origs = new Program.Variable(Origins);
dests = new Program.Variable(Dests);
k = new Program.Variable(SpringKs);
kGround = new Program.Variable(GroundKs);
c = new Program.Variable(Damp);
cGround = new Program.Variable(GroundDamp);
//Outputs
L0 = new Program.Variable(new float[nConnections]);
forces = new Program.Variable(new float[3 * nMasses]);
connForces = new Program.Variable(new float[3 * nConnections]);
nConnec = new Program.Variable(new int[1] { nConnections });
int[] nodesConnects = new int[30 * nMasses];
for (int i = 0; i < nodesConnects.Length; i++) nodesConnects[i] = -1;
nodesConnections = new Program.Variable(nodesConnects);
#endregion
#region Kernels initialization
DEMSource Source = new DEMSource();
string[] s = new string[] { Source.floatcalcL0, Source.floatresetForces, Source.floatcalcForces, Source.floatderivs, Source.floatcalcGroundForces, Source.floatcalcNodesConnections };
CLCalc.Program.Compile(s);
KernelcalcL0 = new Program.Kernel("floatcalcL0");
argscalcL0 = new Program.Variable[] { posOrig, origs, dests, L0 };
KernelcalcNodesConnections = new Program.Kernel("floatcalcNodesConnections");
argscalcNodesConnections = new Program.Variable[] { nodesConnections, nConnec, origs, dests };
KernelresetForces = new Program.Kernel("floatresetForces");
argsresetForces = new Program.Variable[] { forces };
KernelcalcForces = new Program.Kernel("floatcalcForces");
KernelcalcGroundForces = new Program.Kernel("floatcalcGroundForces");
Kernelderivs = new Program.Kernel("floatderivs");
#endregion
// Initial lengths calculation
KernelcalcL0.Execute(argscalcL0, nConn);
//Connections calculation
KernelcalcNodesConnections.Execute(argscalcNodesConnections, nM);
//nodesConnections.ReadFromDeviceTo(nodesConnects);
}
/// <summary>Sums the components of a vector using __local memory and coalesced access</summary>
/// <param name="CLv">Vector whose components should be summed</param>
public static float SumVectorElements(floatVector CLv)
{
float resp = 0;
if (UseOpenCLIfAvailable && CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
/*
The idea here is to create a reduction in which the access pattern to the vectors is coalesced.
The first step is to reduce the number of non-summed items to a multiple of NWORKITEMS and then coalesce the access
*/
int LOCALWORKSIZE = Math.Min(256, (int)CLCalc.Program.CommQueues[CLCalc.Program.DefaultCQ].Device.MaxWorkGroupSize);
int NWORKITEMS = 16 * LOCALWORKSIZE;
int n = CLv.Length;
float[] resps = new float[NWORKITEMS];
if (CLv.CLresps == null)
{
CLv.CLresps = new CLCalc.Program.Variable(resps);
CLv.CLn = new CLCalc.Program.Variable(new int[1]);
}
CLv.CLn.WriteToDevice(new int[] { n });
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { CLv.CLValues, CLv.CLresps, CLv.CLn };
//Write n = k*NWORKITEMS + p. Preprocess to eliminate p`s and leave summation only to a multiple of NWORKITEMS
int k = n / NWORKITEMS;
int p = n - k * NWORKITEMS;
//Clears partial responses
kernelClear.Execute(args, NWORKITEMS);
//Sums the p last elements into the p first elements
kernelPreSum.Execute(args, p);
//Use CLn to inform each work-item its workload. Each one will access and sum k numbers
CLv.CLn.WriteToDevice(new int[] { k });
kernelCoalLocalSum.Execute(args, new int[] { NWORKITEMS }, new int[] { LOCALWORKSIZE });
CLv.CLresps.ReadFromDeviceTo(resps);
//Serial part
int imax = NWORKITEMS / LOCALWORKSIZE;
for (int i = 0; i < imax; i++) resp += resps[i];
}
else
{
double sum = 0;
for (int i = 0; i < CLv.Length; i++) sum += CLv.Values[i];
resp = (float)sum;
}
return resp;
}
/// <summary>Calculates isosurface corresponding to a given isolevel</summary>
/// <param name="isoLvl"></param>
public void CalcIsoSurface(float isoLvl)
{
//Copies iso level to video memory
if (isoLvl != isoLevel[0])
{
isoLevel[0] = isoLvl;
varIsoLevel.WriteToDevice(isoLevel);
}
//Interpolation
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { CLFuncVals, varIsoLevel, varEdgeCoords, varInitVals, varStep };
kernelInterpPts.Execute(args, max);
if (ComputeNormals)
{
//Polygonization
args = new CLCalc.Program.Variable[] { CLFuncVals, varIsoLevel, varEdgeCoords, varEdgePrelimNormals, varElemIndex };
int[] GlobalWorkSize = new int[] { max[0] - 1, max[1] - 1, max[2] - 1 };
kernelPolygonize.Execute(args, GlobalWorkSize);
//Normal smoothing
args = new CLCalc.Program.Variable[] { varEdgePrelimNormals, varEdgeNormals };
kernelSmoothNormals.Execute(args, max);
}
else
{
//Polygonization
args = new CLCalc.Program.Variable[] { CLFuncVals, varIsoLevel, varEdgeCoords, varElemIndex };
int[] GlobalWorkSize = new int[] { max[0] - 1, max[1] - 1, max[2] - 1 };
kernelPolygonizeNoNormals.Execute(args, GlobalWorkSize);
}
}
/// <summary>Initializes physics program. Components indexes: [i] - x, [i+1] - y, [i+2] - z</summary>
/// <param name="nParticles">Number of particles</param>
public floatBodyPhysics(int nParticles)
{
string[] s = new string[] { CollisionAppliers, ForceAppliers, ConstAccelMotionEDOSolver };
Program.Compile(s);
//Kernels
MotionStep = new Program.Kernel("constAccelStep");
Kernel_ApplyGravity = new Program.Kernel("ApplyGravity");
Kernel_FloorCollision = new Program.Kernel("FloorCollision");
Kernel_SelfCollision = new Program.Kernel("SelfCollision");
Kernel_WallCollision = new Program.Kernel("WallCollision");
Kernel_ResetForces = new Program.Kernel("ResetForces");
Kernel_ResetCloseNeighbors = new Program.Kernel("ResetCloseNeighbors");
float[] t = new float[1] { 0 };
float[] gg = new float[3] { 0, 0, 0 };
step = new float[1] { 0 };
//Tamanho de alocacao de velocidades e posicoes
float[] aloc = new float[nParticles * 3];
//Tamanho de alocacao de caracteristicas das particulas
float[] alocPart = new float[nParticles];
//3*Nparticulas
CL_pos = new CLCalc.Program.Variable(aloc);
CL_vel = new CLCalc.Program.Variable(aloc);
CL_forces = new CLCalc.Program.Variable(aloc);
//Nparticulas
closeNeighbors = new int[nParticles];
CL_closeNeighbors = new CLCalc.Program.Variable(closeNeighbors);
for (int i = 0; i < nParticles; i++) alocPart[i] = 1f; //inicializa massas como 1 e tamanhos de colisao como 1
CL_masses = new CLCalc.Program.Variable(alocPart);
CL_collisionSizes = new CLCalc.Program.Variable(alocPart);
//escalares
CL_t = new CLCalc.Program.Variable(t);
CL_step = new CLCalc.Program.Variable(step);
//gravidade
CL_g = new CLCalc.Program.Variable(gg);
//Argumentos de funcoes
stepArgs = new CLCalc.Program.Variable[] { CL_t, CL_step, CL_forces, CL_masses, CL_pos, CL_vel };
applyGravArgs = new CLCalc.Program.Variable[] { CL_forces, CL_masses, CL_g };
floorCollisionArgs = new CLCalc.Program.Variable[] { CL_vel, CL_pos, CL_collisionSizes };
wallCollisionArgs = floorCollisionArgs;
selfCollisionArgs = new CLCalc.Program.Variable[] { CL_vel, CL_pos, CL_masses, CL_forces, CL_closeNeighbors, CL_collisionSizes };
resetForcesArgs = new Program.Variable[] { CL_forces };
resetCloseNeighborsArgs = new Program.Variable[] { CL_closeNeighbors };
nArgs = new int[1] { nParticles * 3 };
nPartics = new int[1] { nParticles };
nPartics2 = new int[2] { nParticles, nParticles };
}
/// <summary>Computes the inverse Discrete Fourier Transform of a float2 vector x whose length is a power of 4.
/// x = { Re[x0] Im[x0] Re[x1] Im[x1] ... Re[xn] Im[xn] }, n = power of 4 (Length = 2*pow(4,n))</summary>
public static CLCalc.Program.Variable iFFT4(CLCalc.Program.Variable CLx)
{
if (CLScale == null) CLScale = new CLCalc.Program.Variable(new float[1]);
//Trick: DFT-1 (x) = DFT(x*)*/N;
//Conjugate
float[] vx = new float[CLx.OriginalVarLength];
CLx.ReadFromDeviceTo(vx);
float[] scale = new float[] { 1 };
CLScale.WriteToDevice(scale);
kernelConjugate.Execute(new CLCalc.Program.Variable[] { CLx, CLScale }, CLx.OriginalVarLength >> 1);
CLx.ReadFromDeviceTo(vx);
CLy = FFT4(ref CLx);
scale[0] = 1 / (float)(CLx.OriginalVarLength >> 1);
CLScale.WriteToDevice(scale);
kernelConjugate.Execute(new CLCalc.Program.Variable[] { CLy, CLScale }, CLy.OriginalVarLength >> 1);
return CLy;
}
/// <summary>Creates the diagonal matrix D(v) with elements associated to those of vector v. Uses the same objects.</summary>
/// <param name="v">Reference vector</param>
public floatDiag(floatVector v)
{
nRows = v.Length;
nCols = v.Length;
this.Values = v.Values;
this.CLValues = v.CLValues;
}
/// <summary>Computes transpose(A)*A and transpose(A)*b weighted by W using OpenCL. Lambda is regularization term</summary>
private static floatSymPosDefMatrix AuxLSAtACL(floatMatrix A, floatDiag W, floatVector lambda, ref floatSymPosDefMatrix AtA)
{
if (AtA == null || AtA.CLValues.OriginalVarLength != (A.Cols * (A.Cols + 1)) >> 1)
{
AtA = new floatSymPosDefMatrix(new float[(A.Cols * (A.Cols + 1)) >> 1]);
}
CLCalc.Program.Variable[] args = new CLCalc.Program.Variable[] { A.CLValues, A.CLDim, W.CLValues, AtA.CLValues, lambda.CLValues };
kernelComputeAtWA.Execute(args, AtA.CLValues.OriginalVarLength);
//Just modified values in CL memory, matrix is no longer Cholesky factorized
AtA.IsCholeskyFactorized = false;
return AtA;
}
public OpenCLTest()
{
CLCalc.InitCL();
var buildlogs = new List<string>();
CLCalc.Program.Compile(new[] { run }, out buildlogs);
Random r = new Random();
float[] sample = Enumerable.Range(0, 768).Select(x => (float)r.NextDouble() * 2 - 1).ToArray();
float[] hlW = Enumerable.Range(0, 10000 * 769).Select(x => (float)r.NextDouble() * 2 - 1).ToArray();
float[] olW = Enumerable.Range(0, 10001 * 10).Select(x => (float)r.NextDouble() * 2 - 1).ToArray();
float[] hlDest = new float[10000];
float[] olDest = new float[10];
//float[] sample = new[] { .3f, .7f };
//float[] hlW = new[] { .4f, .4f, .1f, .5f, .5f, .7f };
//float[] olW = new[] { .4f, .6f, .8f };
//float[] hlDest = new[] { 0f, 0f };
//float[] olDest = new[] { 0f };
var sw = Stopwatch.StartNew();
var sw2 = Stopwatch.StartNew();
var gSample = sample.ToGraphics();
var ghlW = hlW.ToGraphics();
var golW = olW.ToGraphics();
var ghlDest = hlDest.ToGraphics();
var golDest = olDest.ToGraphics();
sw2.Stop();
var kernel = new CLCalc.Program.Kernel("runNN");
var args = new[] {gSample, new CLCalc.Program.Variable(new[] { sample.Length}), ghlDest,
new CLCalc.Program.Variable(new[] { hlDest.Length}), ghlW};
kernel.Execute(args, hlDest.Length);
args = new CLCalc.Program.Variable[5];
args[0] = ghlDest;
args[1] = new CLCalc.Program.Variable(new[] { hlDest.Length });
args[2] = golDest;
args[3] = new CLCalc.Program.Variable(new[] { olDest.Length });
args[4] = golW;
kernel.Execute(args, olDest.Length);
sw2.Start();
ghlDest.ReadFromDeviceTo(hlDest);
golDest.ReadFromDeviceTo(olDest);
sw2.Stop();
sw.Stop();
Console.WriteLine("hlDest\n" + string.Join("\n", hlDest));
Console.WriteLine("olDest\n" + string.Join("\n", olDest));
Console.WriteLine("Total {0}ms", sw.ElapsedMilliseconds);
Console.WriteLine("Memory {0}ms", sw2.ElapsedMilliseconds);
}
/// <summary>Creates a new isosurface calculator. You may pass variables created from a OpenGL context to the CL variables if you are using interop or NULL
/// if not using OpenCL/GL interop.</summary>
/// <param name="FuncValues">Values of the evaluated 3D function f(x,y,z). FuncValues=float[maxX,maxY,maxZ]</param>
/// <param name="CLEdgeCoords">OpenCL variable (float) to hold edge coordinates. Dimension has to be 9 * maxX * maxY * maxZ</param>
/// <param name="CLEdgeNormals">OpenCL variable (float) to hold edge normals. Dimension has to be 9 * maxX * maxY * maxZ</param>
/// <param name="CLElementArrayIndex">OpenCL variable (int) to hold element array index. Dimension has to be 5 * 3 * (maxX - 1) * (maxY - 1) * (maxZ - 1)</param>
private void InitMarchingCubes(float[, ,] FuncValues, CLCalc.Program.Variable CLEdgeCoords, CLCalc.Program.Variable CLEdgeNormals, CLCalc.Program.Variable CLElementArrayIndex)
{
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.Unknown) CLCalc.InitCL();
if (CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
//Reads maximum lengths
int maxX = FuncValues.GetLength(0);
int maxY = FuncValues.GetLength(1);
int maxZ = FuncValues.GetLength(2);
max = new int[] { maxX, maxY, maxZ };
#region Creating variables
//Isolevel
isoLevel = new float[1] { 1.32746E-5f };
varIsoLevel = new CLCalc.Program.Variable(isoLevel);
//Step size and x0,y0,z0
varStep = new CLCalc.Program.Variable(step);
varInitVals = new CLCalc.Program.Variable(initVals);
//Create and copy function values
funcVals = new float[maxX * maxY * maxZ];
CLFuncVals = new CLCalc.Program.Variable(funcVals);
SetFuncVals(FuncValues);
//Edge coordinates - 3 coords * 3 possible directions * number of points
edgeCoords = new float[9 * maxX * maxY * maxZ];
if (CLEdgeCoords != null)
{
varEdgeCoords = CLEdgeCoords;
varEdgeCoords.WriteToDevice(edgeCoords);
}
else varEdgeCoords = new CLCalc.Program.Variable(edgeCoords);
//4 preliminary normals per edge - has to be averaged afterwards
edgePrelimNormals = new float[36 * maxX * maxY * maxZ];
varEdgePrelimNormals = new CLCalc.Program.Variable(edgePrelimNormals);
//Edge normals
edgeNormals = new float[9 * maxX * maxY * maxZ];
if (CLEdgeNormals != null)
{
varEdgeNormals = CLEdgeNormals;
varEdgeNormals.WriteToDevice(edgeNormals);
}
else varEdgeNormals = new CLCalc.Program.Variable(edgeNormals);
//Number of cubes: (maxX-1)*(maxY-1)*(maxZ-1)
//Marching cube algorithm: each cube can have 5 triangles drawn, 3 vertexes per triangle
//q-th vertex of p-th triangle of the ijk-th cube: [(5*(i+(maxX-1)*j+k*(maxX-1)*(maxY-1))+p)*3+q]
elementIndex = new int[5 * 3 * (maxX - 1) * (maxY - 1) * (maxZ - 1)];
if (CLElementArrayIndex != null)
{
varElemIndex = CLElementArrayIndex;
varElemIndex.WriteToDevice(elementIndex);
}
else varElemIndex = new CLCalc.Program.Variable(elementIndex);
//Edge remapping to build output
edges = new int[edgeCoords.Length / 3];
for (int i = 0; i < edges.Length; i++) edges[i] = -1;
#endregion
#region Compile code and create kernels
CLMarchingCubesSrc cmsrc = new CLMarchingCubesSrc();
CLCalc.Program.Compile(new string[] { cmsrc.definitions, cmsrc.src });
kernelInterpPts = new CLCalc.Program.Kernel("interpPts");
kernelPolygonize = new CLCalc.Program.Kernel("Polygonize");
kernelSmoothNormals = new CLCalc.Program.Kernel("SmoothNormals");
kernelPolygonizeNoNormals = new CLCalc.Program.Kernel("PolygonizeNoNormals");
#endregion
}
else throw new Exception("OpenCL not available");
}
/// <summary>OpenCL diagonal matrix constructor</summary>
/// <param name="Vals">Main diagonal elements</param>
public floatDiag(float[] Vals)
{
this.Values = (float[])Vals.Clone();
if (UseOpenCLIfAvailable && CLCalc.CLAcceleration == CLCalc.CLAccelerationType.UsingCL)
{
CLValues = new CLCalc.Program.Variable(Values);
}
nRows = Vals.Length;
nCols = Vals.Length;
}
