private void DoExecute(FFTPlan plan, object input, object output, bool inverse = false)
{
FFTPlanEx planEx = Plans[plan];
CUDevicePtrEx inPtrEx;
CUDevicePtrEx outPtrEx;
inPtrEx = _gpu.GetDeviceMemory(input) as CUDevicePtrEx;
outPtrEx = _gpu.GetDeviceMemory(output) as CUDevicePtrEx;
CUFFTDirection dir = inverse ? CUFFTDirection.Inverse : CUFFTDirection.Forward;
CUFFTResult res = CUFFTResult.ExecFailed;
if (planEx.CudaFFTType == CUFFTType.C2C)
{
res = _driver.cufftExecC2C(planEx.CudaFFTHandle, inPtrEx.DevPtr, outPtrEx.DevPtr, dir);
}
else if (planEx.CudaFFTType == CUFFTType.C2R)
{
res = _driver.cufftExecC2R(planEx.CudaFFTHandle, inPtrEx.DevPtr, outPtrEx.DevPtr);
}
else if (planEx.CudaFFTType == CUFFTType.D2Z)
{
res = _driver.cufftExecD2Z(planEx.CudaFFTHandle, inPtrEx.DevPtr, outPtrEx.DevPtr);
}
else if (planEx.CudaFFTType == CUFFTType.R2C)
{
res = _driver.cufftExecR2C(planEx.CudaFFTHandle, inPtrEx.DevPtr, outPtrEx.DevPtr);
}
else if (planEx.CudaFFTType == CUFFTType.Z2D)
{
res = _driver.cufftExecZ2D(planEx.CudaFFTHandle, inPtrEx.DevPtr, outPtrEx.DevPtr);
}
else if (planEx.CudaFFTType == CUFFTType.Z2Z)
{
res = _driver.cufftExecZ2Z(planEx.CudaFFTHandle, inPtrEx.DevPtr, outPtrEx.DevPtr, dir);
}
if (res != CUFFTResult.Success)
{
throw new CudafyMathException(res.ToString());
}
}
public void ExecuteComplexToComplex(cufftHandle plan, CUdeviceptr input, CUdeviceptr output, CUFFTDirection direction)
{
this.LastError = _driver.cufftExecC2C(plan, input, output, direction);
}
//public void Execute1D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int batch)
//{
// this.Execute1D(input, output, nx, batch, CUFFTDirection.Forward);
//}
//public void Execute1D(cuFloatComplex[] input, cuFloatReal[] output, int nx, int batch)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan1d(ref plan, nx, CUFFTType.C2R, batch);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatReal>(output);
// CUFFTDriver.cufftExecC2R(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatReal>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute1D(cuFloatReal[] input, cuFloatComplex[] output, int nx, int batch)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan1d(ref plan, nx, CUFFTType.R2C, batch);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatReal>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecR2C(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute1D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int batch, CUFFTDirection direction)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan1d(ref plan, nx, CUFFTType.C2C, batch);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecC2C(plan, idata, odata, direction);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute2D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny)
//{
// this.Execute2D(input, output, nx, ny, CUFFTDirection.Forward);
//}
//public void Execute2D(cuFloatComplex[] input, cuFloatReal[] output, int nx, int ny)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan2d(ref plan, nx, ny, CUFFTType.C2R);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatReal>(output);
// CUFFTDriver.cufftExecC2R(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatReal>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute2D(cuFloatReal[] input, cuFloatComplex[] output, int nx, int ny)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan2d(ref plan, nx, ny, CUFFTType.R2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatReal>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecR2C(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute2D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny, CUFFTDirection direction)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan2d(ref plan, nx, ny, CUFFTType.C2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecC2C(plan, idata, odata, direction);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute3D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny, int nz)
//{
// this.Execute3D(input, output, nx, ny, nz, CUFFTDirection.Forward);
//}
//public void Execute3D(cuFloatComplex[] input, cuFloatReal[] output, int nx, int ny, int nz)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan3d(ref plan, nx, ny, nz, CUFFTType.C2R);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatReal>(output);
// CUFFTDriver.cufftExecC2R(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatReal>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute3D(cuFloatReal[] input, cuFloatComplex[] output, int nx, int ny, int nz)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan3d(ref plan, nx, ny, nz, CUFFTType.R2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatReal>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecR2C(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute3D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny, int nz, CUFFTDirection direction)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan3d(ref plan, nx, ny, nz, CUFFTType.C2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecC2C(plan, idata, odata, direction);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
public void ExecuteComplexToComplex(CUdeviceptr input, CUdeviceptr output, CUFFTDirection direction)
{
this.ExecuteComplexToComplex(this.plan, input, output, direction);
}
public CUFFTResult cufftExecZ2Z(cufftHandle plan, CUdeviceptr idata, CUdeviceptr odata, CUFFTDirection direction)
{
return(cufftExecZ2Z_ext(plan, idata, odata, direction));
}
private static extern CUFFTResult cufftExecZ2Z_ext(cufftHandle plan, CUdeviceptr idata, CUdeviceptr odata, CUFFTDirection direction);
public static extern CUFFTResult cufftExecZ2Z(cufftHandle plan, CUdeviceptr idata, CUdeviceptr odata, CUFFTDirection direction);
public CUFFTResult cufftExecZ2Z(cufftHandle plan, CUdeviceptr idata, CUdeviceptr odata, CUFFTDirection direction)
{
return cufftExecZ2Z_ext(plan, idata, odata, direction);
}
private static extern CUFFTResult cufftExecZ2Z_ext(cufftHandle plan, CUdeviceptr idata, CUdeviceptr odata, CUFFTDirection direction);
public void ExecuteComplexToComplex(cufftHandle plan, CUdeviceptr input, CUdeviceptr output, CUFFTDirection direction)
{
this.LastError = _driver.cufftExecC2C(plan, input, output, direction);
}
//public void Execute1D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int batch)
//{
// this.Execute1D(input, output, nx, batch, CUFFTDirection.Forward);
//}
//public void Execute1D(cuFloatComplex[] input, cuFloatReal[] output, int nx, int batch)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan1d(ref plan, nx, CUFFTType.C2R, batch);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatReal>(output);
// CUFFTDriver.cufftExecC2R(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatReal>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute1D(cuFloatReal[] input, cuFloatComplex[] output, int nx, int batch)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan1d(ref plan, nx, CUFFTType.R2C, batch);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatReal>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecR2C(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute1D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int batch, CUFFTDirection direction)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan1d(ref plan, nx, CUFFTType.C2C, batch);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecC2C(plan, idata, odata, direction);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute2D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny)
//{
// this.Execute2D(input, output, nx, ny, CUFFTDirection.Forward);
//}
//public void Execute2D(cuFloatComplex[] input, cuFloatReal[] output, int nx, int ny)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan2d(ref plan, nx, ny, CUFFTType.C2R);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatReal>(output);
// CUFFTDriver.cufftExecC2R(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatReal>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute2D(cuFloatReal[] input, cuFloatComplex[] output, int nx, int ny)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan2d(ref plan, nx, ny, CUFFTType.R2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatReal>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecR2C(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute2D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny, CUFFTDirection direction)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan2d(ref plan, nx, ny, CUFFTType.C2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecC2C(plan, idata, odata, direction);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute3D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny, int nz)
//{
// this.Execute3D(input, output, nx, ny, nz, CUFFTDirection.Forward);
//}
//public void Execute3D(cuFloatComplex[] input, cuFloatReal[] output, int nx, int ny, int nz)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan3d(ref plan, nx, ny, nz, CUFFTType.C2R);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatReal>(output);
// CUFFTDriver.cufftExecC2R(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatReal>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute3D(cuFloatReal[] input, cuFloatComplex[] output, int nx, int ny, int nz)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan3d(ref plan, nx, ny, nz, CUFFTType.R2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatReal>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecR2C(plan, idata, odata);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
//public void Execute3D(cuFloatComplex[] input, cuFloatComplex[] output, int nx, int ny, int nz, CUFFTDirection direction)
//{
// cufftHandle plan = new cufftHandle();
// this.LastError = CUFFTDriver.cufftPlan3d(ref plan, nx, ny, nz, CUFFTType.C2C);
// CUdeviceptr idata = this.cuda.CopyHostToDevice<cuFloatComplex>(input);
// CUdeviceptr odata = this.cuda.Allocate<cuFloatComplex>(output);
// CUFFTDriver.cufftExecC2C(plan, idata, odata, direction);
// this.cuda.CopyDeviceToHost<cuFloatComplex>(odata, output);
// CUFFTDriver.cufftDestroy(plan);
// this.cuda.Free(idata);
// this.cuda.Free(odata);
//}
public void ExecuteComplexToComplex(CUdeviceptr input, CUdeviceptr output, CUFFTDirection direction)
{
this.ExecuteComplexToComplex(this.plan, input, output, direction);
}
public static extern CUFFTResult cufftExecZ2Z(cufftHandle plan, CUdeviceptr idata, CUdeviceptr odata, CUFFTDirection direction);