/// <summary>
/// Configures the layout of CUFFT output in FFTW‐compatible modes.
/// When FFTW compatibility is desired, it can be configured for padding
/// only, for asymmetric complex inputs only, or to be fully compatible.
/// </summary>
/// <param name="plan">The plan.</param>
/// <param name="mode">The mode.</param>
public override void SetCompatibilityMode(FFTPlan plan, eCompatibilityMode mode)
{
CUFFTCompatibility cumode = (CUFFTCompatibility)mode;
FFTPlanEx planEx = Plans[plan];
CUFFTResult res = _driver.cufftSetCompatibilityMode(planEx.CudaFFTHandle, cumode);
if (res != CUFFTResult.Success)
{
throw new CudafyHostException(res.ToString());
}
}
private void DoExecute <T, U>(FFTPlan plan, Array input, Array output, bool inverse = false)
{
//_gpu.VerifyOnGPU(input);
//_gpu.VerifyOnGPU(output);
EmuDevicePtrEx inputPtr = _gpu.GetDeviceMemory(input) as EmuDevicePtrEx;
EmuDevicePtrEx outputPtr = _gpu.GetDeviceMemory(output) as EmuDevicePtrEx;
FFTPlanEx planEx = Plans[plan];
Ifftw_plan fftwplan = inverse ? planEx.FFTWInvPlan : planEx.FFTWFwdPlan;
int insize = Marshal.SizeOf(typeof(T));
int inoffset = inputPtr.OffsetBytes;
int outoffset = outputPtr.OffsetBytes;
int outsize = Marshal.SizeOf(typeof(U));
int batchSize = plan.BatchSize;
int planLength = plan.Length;
int N = planEx.N;
unsafe
{
GCHandle inhandle = new GCHandle();
GCHandle outhandle = new GCHandle();
try
{
inhandle = GCHandle.Alloc(inputPtr.DevPtr, GCHandleType.Pinned);
outhandle = GCHandle.Alloc(outputPtr.DevPtr, GCHandleType.Pinned);
long srcAddress = inhandle.AddrOfPinnedObject().ToInt64();
long dstAddress = outhandle.AddrOfPinnedObject().ToInt64();
int srcOffset = 0;
int dstOffset = 0;
//for (int b = 0; b < batchSize; b++)
{
IntPtr srcOffsetPtr = new IntPtr(srcAddress + inoffset + (srcOffset * insize));
GPGPU.CopyMemory(fftwplan.Input, srcOffsetPtr, (uint)(N * insize * batchSize));
fftwplan.Execute();
IntPtr dstIntPtrOffset = new IntPtr(dstAddress + outoffset + (dstOffset * outsize));
GPGPU.CopyMemory(dstIntPtrOffset, fftwplan.Output, (uint)(N * outsize * batchSize));
//srcOffset += planLength;
// dstOffset += planLength;
}
}
finally
{
inhandle.Free();
outhandle.Free();
}
}
}
//[DllImport("kernel32.dll", EntryPoint = "RtlMoveMemory")]
//private static extern void CopyMemory(IntPtr Destination, IntPtr Source, uint Length);
/// <summary>
/// Frees the specified plan.
/// </summary>
/// <param name="plan">The plan.</param>
public override void Remove(FFTPlan plan)
{
FFTPlanEx planEx = Plans[plan];
if (planEx.DataType == eDataType.Single)
{
fftwf.destroy_plan(planEx.FFTWFwdPlan.Handle);
fftwf.destroy_plan(planEx.FFTWInvPlan.Handle);
}
else if (planEx.DataType == eDataType.Double)
{
fftw.destroy_plan(planEx.FFTWFwdPlan.Handle);
fftw.destroy_plan(planEx.FFTWInvPlan.Handle);
}
Plans.Remove(plan);
}
/// <summary>
/// Sets the stream.
/// </summary>
/// <param name="plan">The plan to set the stream for.</param>
/// <param name="streamId">The stream id.</param>
public override void SetStream(FFTPlan plan, int streamId)
{
if (streamId < 0)
{
throw new ArgumentOutOfRangeException("streamId");
}
CUstream cus = (CUstream)_gpu.GetStream(streamId);
FFTPlanEx planEx = Plans[plan];
cudaStream cs = new cudaStream();
//cs.Value = cus.Pointer.ToInt32();
CUFFTResult res = _driver.cufftSetStream(planEx.CudaFFTHandle, cs);
if (res != CUFFTResult.Success)
{
throw new CudafyMathException(CudafyMathException.csCUDA_EXCEPTION_X, res);
}
}
/// <summary>
/// Frees the specified plan.
/// </summary>
/// <param name="plan">The plan.</param>
public override void Remove(FFTPlan plan)
{
FFTPlanEx planEx = Plans[plan];
CUFFTResult res = _driver.cufftDestroy(planEx.CudaFFTHandle);
if (res != CUFFTResult.Success)
{
//throw new CudafyHostException(res.ToString());
Debug.WriteLine("remove plan failed: " + res.ToString());
}
else
{
Debug.WriteLine("remove plan succeeded: " + res.ToString());
}
Plans.Remove(plan);
}
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());
}
}