public MyFourierBinder(MyWorkingNode owner, int inputSize, MyMemoryBlock <float> tempBlock)
: base(owner, inputSize, tempBlock)
{
m_fft = new CudaFFTPlan1D(inputSize, cufftType.R2C, 1);
m_ifft = new CudaFFTPlan1D(inputSize, cufftType.C2R, 1);
m_mulkernel = MyKernelFactory.Instance.Kernel(owner.GPU, @"Common\CombineVectorsKernel", "MulComplexElementWise");
m_mulkernel.SetupExecution(inputSize + 1);
m_involutionKernel = MyKernelFactory.Instance.Kernel(owner.GPU, @"Common\CombineVectorsKernel", "InvolveVector");
m_involutionKernel.SetupExecution(inputSize - 1);
m_inversionKernel = MyKernelFactory.Instance.Kernel(owner.GPU, @"Transforms\InvertValuesKernel", "InvertLengthComplexKernel");
m_inversionKernel.SetupExecution(inputSize);
m_dotKernel = MyReductionFactory.Kernel(owner.GPU, MyReductionFactory.Mode.f_DotProduct_f);
m_normalKernel = MyKernelFactory.Instance.Kernel(owner.GPU, @"Transforms\TransformKernels", "PolynomialFunctionKernel");
m_normalKernel.SetupExecution(inputSize);
m_firstFFTOffset = 0;
m_secondFFTOffset = (inputSize + 1) * 2;
m_tempOffset = (inputSize + 1) * 4;
Denominator = inputSize;
}
public List <float> hypotesis(List <double> x, List <double> h, int N)
{
//int N = 2000000;
string path = Path.GetDirectoryName(mv.plugins[0].filename);
CudaContext ctx = new CudaContext();
CudaKernel kernel = ctx.LoadKernel(path + "\\kernel.ptx", "ComplexMultCUDA");
kernel.GridDimensions = (int)Math.Ceiling((double)(N + h.Count - 1) / 1024);
kernel.BlockDimensions = 1024;
double[] temp_y = new double[N + h.Count - 1];
double[] temp_h = new double[N + h.Count - 1];
double[] temp_x = new double[N + h.Count - 1];
double2[] temp_x2 = new double2[N + h.Count - 1];
h.ToArray().CopyTo(temp_h, 0);
x.ToArray().CopyTo(temp_x, 0);
CudaDeviceVariable <double> d_x = null;
CudaDeviceVariable <double2> d_X = new CudaDeviceVariable <double2>(N + h.Count - 1);
CudaDeviceVariable <double> d_h = new CudaDeviceVariable <double>(N + h.Count - 1);
CudaDeviceVariable <double2> d_H = new CudaDeviceVariable <double2>(N + h.Count - 1);
CudaDeviceVariable <double> d_y = new CudaDeviceVariable <double>(N + h.Count - 1);
CudaFFTPlan1D planForward = new CudaFFTPlan1D(N + h.Count - 1, cufftType.D2Z, 1);
CudaFFTPlan1D planInverse = new CudaFFTPlan1D(N + h.Count - 1, cufftType.Z2D, 1);
try
{
d_h = temp_h;
planForward.Exec(d_h.DevicePointer, d_H.DevicePointer, TransformDirection.Forward);
}
catch (Exception exp)
{
mainView.log(exp, "CUDA error: Impulse response FFT", this);
return(null);
}
try
{
d_x = temp_x;
planForward.Exec(d_x.DevicePointer, d_X.DevicePointer);
kernel.Run(d_H.DevicePointer, d_X.DevicePointer, N + h.Count - 1);
planInverse.Exec(d_X.DevicePointer, d_y.DevicePointer);
}
catch (Exception exp)
{
mainView.log(exp, "Cuda error: kernel run cuda error", this);
}
temp_y = d_y;
return(Array.ConvertAll <double, float>(temp_y, d => (float)d).ToList().GetRange(500, x.Count));
}
public List <float> CUDA_FIR(List <float> x, List <double> h)
{
CudaContext ctx = new CudaContext();
//alloc data to cuda format
double2[] temp_x = new double2[x.Count + h.Count - 1];
double2[] temp_h = new double2[x.Count + h.Count - 1];
double2[] temp_y = new double2[x.Count + h.Count - 1];
//data copy
for (int i = 0; i < x.Count; i++)
{
temp_x[i].x = x[i];
}
for (int i = 0; i < h.Count; i++)
{
temp_h[i].x = h[i];
}
CudaDeviceVariable <double2> d_x = null;
CudaDeviceVariable <double2> d_h = null;
CudaFFTPlan1D plan1D = new CudaFFTPlan1D(x.Count + h.Count - 1, cufftType.Z2Z, 1);
CudaKernel kernel = ctx.LoadKernel("kernel.ptx", "ComplexMultCUDA");
kernel.GridDimensions = (int)Math.Ceiling((double)(x.Count + h.Count - 1) / 1024);
kernel.BlockDimensions = 1024;
try
{
d_x = temp_x;
d_h = temp_h;
}
catch (Exception e)
{
//("{0} Exception caught.", e);
return(null);
}
plan1D.Exec(d_x.DevicePointer, TransformDirection.Forward);
plan1D.Exec(d_h.DevicePointer, TransformDirection.Forward);
kernel.Run(d_h.DevicePointer, d_x.DevicePointer, x.Count + h.Count - 1);
plan1D.Exec(d_x.DevicePointer, TransformDirection.Inverse);
temp_y = d_x;
return(temp_y.Select(data => (float)data.x).ToList().GetRange(h.Count / 2, x.Count));
}
public override void Init(int nGPU)
{
fft = new CudaFFTPlan1D(Owner.InputSize, cufftType.R2C, 1);
ifft = new CudaFFTPlan1D(Owner.InputSize, cufftType.C2R, 1);
m_kernel = MyKernelFactory.Instance.Kernel(nGPU, @"Common\CombineVectorsKernel", "MulComplexElementWise");
m_kernel.SetupExecution(Owner.InputSize + 1);
m_involutionKernel = MyKernelFactory.Instance.Kernel(nGPU, @"Common\CombineVectorsKernel", "InvolveVector");
m_involutionKernel.SetupExecution(Owner.InputSize - 1);
m_linearCombKernel = MyKernelFactory.Instance.Kernel(nGPU, @"Transforms\TransformKernels", "LinearCombinationKernel");
m_linearCombKernel.SetupExecution(Owner.InputSize);
m_normalKernel = MyKernelFactory.Instance.Kernel(nGPU, @"Transforms\TransformKernels", "PolynomialFunctionKernel");
m_normalKernel.SetupExecution(Owner.InputSize);
}
public void cuFFTreconstruct()
{
CudaContext ctx = new CudaContext(0);
ManagedCuda.BasicTypes.CUmodule cumodule = ctx.LoadModule("kernel.ptx");
CudaKernel cuKernel = new CudaKernel("cu_ArrayInversion", cumodule, ctx);
float2[] fData = new float2[Resolution * Resolution];
float2[] result = new float2[Resolution * Resolution];
FFTData2D = new float[Resolution, Resolution, 2];
CudaDeviceVariable <float2> devData = new CudaDeviceVariable <float2>(Resolution * Resolution);
CudaDeviceVariable <float2> copy_devData = new CudaDeviceVariable <float2>(Resolution * Resolution);
int i, j;
Random rnd = new Random();
double avrg = 0.0;
for (i = 0; i < Resolution; i++)
{
for (j = 0; j < Resolution; j++)
{
fData[i * Resolution + j].x = i + j * 2;
avrg += fData[i * Resolution + j].x;
fData[i * Resolution + j].y = 0.0f;
}
}
avrg = avrg / (double)(Resolution * Resolution);
for (i = 0; i < Resolution; i++)
{
for (j = 0; j < Resolution; j++)
{
fData[(i * Resolution + j)].x = fData[(i * Resolution + j)].x - (float)avrg;
}
}
devData.CopyToDevice(fData);
CudaFFTPlan1D plan1D = new CudaFFTPlan1D(Resolution, cufftType.C2C, Resolution);
plan1D.Exec(devData.DevicePointer, TransformDirection.Forward);
cuKernel.GridDimensions = new ManagedCuda.VectorTypes.dim3(Resolution / cuda_blockNum, Resolution, 1);
cuKernel.BlockDimensions = new ManagedCuda.VectorTypes.dim3(cuda_blockNum, 1, 1);
cuKernel.Run(devData.DevicePointer, copy_devData.DevicePointer, Resolution);
copy_devData.CopyToHost(result);
for (i = 0; i < Resolution; i++)
{
for (j = 0; j < Resolution; j++)
{
FFTData2D[i, j, 0] = result[i * Resolution + j].x;
FFTData2D[i, j, 1] = result[i * Resolution + j].y;
}
}
//Clean up
devData.Dispose();
copy_devData.Dispose();
plan1D.Dispose();
CudaContext.ProfilerStop();
ctx.Dispose();
}
static void Main(string[] args)
{
int SIGNAL_SIZE = 50;
int FILTER_KERNEL_SIZE = 11;
Console.WriteLine("[simpleCUFFT] is starting...");
var assembly = Assembly.GetExecutingAssembly();
var resourceName = "simpleCUFFT.simpleCUFFTKernel.ptx";
CudaContext ctx = new CudaContext(0);
CudaKernel ComplexPointwiseMulAndScale;
string[] liste = assembly.GetManifestResourceNames();
using (Stream stream = assembly.GetManifestResourceStream(resourceName))
{
ComplexPointwiseMulAndScale = ctx.LoadKernelPTX(stream, "ComplexPointwiseMulAndScale");
}
// Allocate host memory for the signal
cuFloatComplex[] h_signal = new cuFloatComplex[SIGNAL_SIZE]; //we use cuFloatComplex for complex multiplaction in reference host code...
Random rand = new Random(0);
// Initialize the memory for the signal
for (int i = 0; i < SIGNAL_SIZE; ++i)
{
h_signal[i].real = (float)rand.NextDouble();
h_signal[i].imag = 0;
}
// Allocate host memory for the filter
cuFloatComplex[] h_filter_kernel = new cuFloatComplex[FILTER_KERNEL_SIZE];
// Initialize the memory for the filter
for (int i = 0; i < FILTER_KERNEL_SIZE; ++i)
{
h_filter_kernel[i].real = (float)rand.NextDouble();
h_filter_kernel[i].imag = 0;
}
// Pad signal and filter kernel
cuFloatComplex[] h_padded_signal = null;
cuFloatComplex[] h_padded_filter_kernel = null;
int new_size = PadData(h_signal, ref h_padded_signal, SIGNAL_SIZE,
h_filter_kernel, ref h_padded_filter_kernel, FILTER_KERNEL_SIZE);
int mem_size = (int)cuFloatComplex.SizeOf * new_size;
// Allocate device memory for signal
CudaDeviceVariable <cuFloatComplex> d_signal = new CudaDeviceVariable <cuFloatComplex>(new_size);
// Copy host memory to device
d_signal.CopyToDevice(h_padded_signal);
// Allocate device memory for filter kernel
CudaDeviceVariable <cuFloatComplex> d_filter_kernel = new CudaDeviceVariable <cuFloatComplex>(new_size);
// Copy host memory to device
d_filter_kernel.CopyToDevice(h_padded_filter_kernel);
// CUFFT plan simple API
CudaFFTPlan1D plan = new CudaFFTPlan1D(new_size, cufftType.C2C, 1);
// Transform signal and kernel
Console.WriteLine("Transforming signal cufftExecC2C");
plan.Exec(d_signal.DevicePointer, TransformDirection.Forward);
plan.Exec(d_filter_kernel.DevicePointer, TransformDirection.Forward);
// Multiply the coefficients together and normalize the result
Console.WriteLine("Launching ComplexPointwiseMulAndScale<<< >>>");
ComplexPointwiseMulAndScale.BlockDimensions = 256;
ComplexPointwiseMulAndScale.GridDimensions = 32;
ComplexPointwiseMulAndScale.Run(d_signal.DevicePointer, d_filter_kernel.DevicePointer, new_size, 1.0f / new_size);
// Transform signal back
Console.WriteLine("Transforming signal back cufftExecC2C");
plan.Exec(d_signal.DevicePointer, TransformDirection.Inverse);
// Copy device memory to host
cuFloatComplex[] h_convolved_signal = d_signal;
// Allocate host memory for the convolution result
cuFloatComplex[] h_convolved_signal_ref = new cuFloatComplex[SIGNAL_SIZE];
// Convolve on the host
Convolve(h_signal, SIGNAL_SIZE,
h_filter_kernel, FILTER_KERNEL_SIZE,
h_convolved_signal_ref);
// check result
bool bTestResult = sdkCompareL2fe(h_convolved_signal_ref, h_convolved_signal, 1e-5f);
//Destroy CUFFT context
plan.Dispose();
// cleanup memory
d_filter_kernel.Dispose();
d_signal.Dispose();
ctx.Dispose();
if (bTestResult)
{
Console.WriteLine("Test Passed");
}
else
{
Console.WriteLine("Test Failed");
}
}
public List <float> hypotesis_long_save(List <double> xx, List <double> h, int N)
{
int n = (int)Math.Ceiling((double)(xx.Count() + 0.000000000001) / N);
double[] temp_data = new double[n * (N + h.Count - 1) - (n - 1) * (h.Count - 1)];
xx.CopyTo(temp_data, h.Count - 1);
List <double> x = temp_data.ToList();
//int N = 2000000;
string path = Path.GetDirectoryName(mv.plugins[0].filename);
CudaContext ctx = new CudaContext();
CudaKernel kernel = ctx.LoadKernel(path + "\\kernel.ptx", "ComplexMultCUDA");
kernel.GridDimensions = (int)Math.Ceiling((double)(N + h.Count - 1) / 1024);
kernel.BlockDimensions = 1024;
int blocks = (int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1));
double[][] temp_y = new double[n][];
double[] temp_h = new double[N + h.Count - 1];
double[] temp_x = new double[N + h.Count - 1];
h.ToArray().CopyTo(temp_h, 0);
CudaDeviceVariable <double> d_x = null;
CudaDeviceVariable <double> d_h = new CudaDeviceVariable <double>(N + h.Count - 1);
CudaDeviceVariable <double2> d_H = new CudaDeviceVariable <double2>(N + h.Count - 1);
//CudaDeviceVariable<double> d_y = new CudaDeviceVariable<double>(N + h.Count - 1);
CudaFFTPlan1D planForward = new CudaFFTPlan1D(N + h.Count - 1, cufftType.D2Z, 1);
CudaFFTPlan1D planInverse = new CudaFFTPlan1D(N + h.Count - 1, cufftType.Z2D, 1);
try
{
d_h = temp_h;
planForward.Exec(d_h.DevicePointer, d_H.DevicePointer, TransformDirection.Forward);
}
catch (Exception exp)
{
mainView.log(exp, "CUDA error: Impulse response FFT", this);
return(null);
}
for (int g = 0; g < n; g++)
{
CudaDeviceVariable <double2> d_X = new CudaDeviceVariable <double2>(N + h.Count - 1);
int P = N + h.Count - 1;
//if (x.Count - P * g < P) P = x.Count - P * g;
int L = h.Count - 1;
if (g == 0)
{
L = 0;
}
x.CopyTo(P * g - L * g, temp_x, 0, P);
try
{
d_x = temp_x;
planForward.Exec(d_x.DevicePointer, d_X.DevicePointer);
kernel.Run(d_H.DevicePointer, d_X.DevicePointer, N + h.Count - 1);
planInverse.Exec(d_X.DevicePointer, d_x.DevicePointer);
}
catch (Exception exp)
{
mainView.log(exp, "Cuda error: kernel run cuda error", this);
}
temp_y[g] = d_x;
d_x.Dispose();
d_X.Dispose();
}
planForward.Dispose();
planInverse.Dispose();
d_x.Dispose();
d_h.Dispose();
d_H.Dispose();
ctx.Dispose();
return(OverlapSave(temp_y, h.Count, N + h.Count - 1).GetRange(h.Count / 2, xx.Count));
}
public List <float> hypotesis_long(List <double> x, List <double> h, int N)
{
//int N = 2000000;
string path = Path.GetDirectoryName(mv.plugins[0].filename);
CudaContext ctx = new CudaContext();
CudaKernel kernel = ctx.LoadKernel(path + "\\kernel.ptx", "ComplexMultCUDA");
kernel.GridDimensions = (int)Math.Ceiling((double)(N + h.Count - 1) / 1024);
kernel.BlockDimensions = 1024;
int blocks = (int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1));
double[][] temp_y = new double[blocks][];
double[] temp_h = new double[N + h.Count - 1];
double[] temp_x = new double[N + h.Count - 1];
h.ToArray().CopyTo(temp_h, 0);
CudaDeviceVariable <double> d_x = null;
CudaDeviceVariable <double2> d_X = new CudaDeviceVariable <double2>(N + h.Count - 1);
CudaDeviceVariable <double> d_h = new CudaDeviceVariable <double>(N + h.Count - 1);
CudaDeviceVariable <double2> d_H = new CudaDeviceVariable <double2>(N + h.Count - 1);
//CudaDeviceVariable<double> d_y = new CudaDeviceVariable<double>(N + h.Count - 1);
CudaFFTPlan1D planForward = new CudaFFTPlan1D(N + h.Count - 1, cufftType.D2Z, 1);
CudaFFTPlan1D planInverse = new CudaFFTPlan1D(N + h.Count - 1, cufftType.Z2D, 1);
try
{
d_h = temp_h;
planForward.Exec(d_h.DevicePointer, d_H.DevicePointer, TransformDirection.Forward);
}
catch (Exception exp)
{
mainView.log(exp, "CUDA error: Impulse response FFT", this);
return(null);
}
for (int g = 0; g < blocks; g++)
{
int P = N;
if (x.Count - N * g < N)
{
P = x.Count - N * g;
}
x.GetRange(N * g, P).ToArray().CopyTo(temp_x, 0);
try
{
d_x = temp_x;
planForward.Exec(d_x.DevicePointer, d_X.DevicePointer);
kernel.Run(d_H.DevicePointer, d_X.DevicePointer, N + h.Count - 1);
planInverse.Exec(d_X.DevicePointer, d_x.DevicePointer);
}
catch (Exception exp)
{
mainView.log(exp, "Cuda error: kernel run cuda error", this);
}
temp_y[g] = d_x;
}
return(OverlapAdd(temp_y, h.Count).GetRange(h.Count / 2, x.Count));
}
public List <float> CUDA_FIR_long(List <float> x, List <double> h)
{
CudaContext ctx = new CudaContext();
string path = Path.GetDirectoryName(mv.plugins[0].filename);
int N = 2000000;
//alloc data to cuda format
double2[][] temp_x = new double2[(int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1))][];
double2[] temp_h = new double2[N + h.Count - 1];
double2[][] temp_y = new double2[(int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1))][];
//data copy
System.Threading.Tasks.Parallel.For(0, (int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1)), j => {
temp_x[j] = new double2[N + h.Count - 1];
temp_y[j] = new double2[N + h.Count - 1];
for (int i = 0; (j * N + i) < x.Count && i < N; i++)
{
temp_x[j][i].x = x[j * N + i];
}
});
for (int i = 0; i < h.Count; i++)
{
temp_h[i].x = h[i];
}
CudaDeviceVariable <double2> d_x = null;
CudaDeviceVariable <double2> d_h = null;
CudaFFTPlan1D plan1D = new CudaFFTPlan1D(N + h.Count - 1, cufftType.Z2Z, 1);
CudaKernel kernel = ctx.LoadKernel(path + "\\kernel.ptx", "ComplexMultCUDA");
kernel.GridDimensions = (int)Math.Ceiling((double)(N + h.Count - 1) / 1024);
kernel.BlockDimensions = 1024;
try
{
d_h = temp_h;
}
catch (Exception e)
{
//("{0} Exception caught.", e);
return(null);
}
plan1D.Exec(d_h.DevicePointer, TransformDirection.Forward);
for (int g = 0; g < (int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1)); g++)
{
try
{
d_x = temp_x[g];
}
catch (Exception e)
{
mainView.log(e, "cuda alloc data error", this);
return(null);
}
try
{
plan1D.Exec(d_x.DevicePointer, TransformDirection.Forward);
kernel.Run(d_h.DevicePointer, d_x.DevicePointer, N + h.Count - 1);
plan1D.Exec(d_x.DevicePointer, TransformDirection.Inverse);
}
catch (Exception exp)
{
mainView.log(exp, "kernel run cuda error", this);
}
temp_y[g] = d_x;
//this.Invoke((MethodInvoker)delegate
//{
// progressBar1.Value = (int)(50/ (int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1)))*g;
//});
d_x.Dispose();
}
d_h.Dispose();
plan1D.Dispose();
return(OverlapAdd(temp_y, h.Count).GetRange(h.Count / 2, x.Count));
}
