DFTI performance improvement

[TobiasErnst]

In the current implementation of MKL.NET a descriptor is created for every DFT.
If you need to run similar DFTs with only different data this results in a not needed overhead due to memory allocations.

Therefore the methods DftiCreateDescriptor and DftiFreeDescriptor should also be accessible via MKL.NET and it should be possible to inject an already created descriptor to all other Dfti methods.

I am willing to implement this but I currently have a problem with the Package-Native.ps1 whish throws the error

Exception calling "DownloadFile" with "2" argument(s): "The remote server returned an error: (403) Forbidden."
At D:\MKL.NET\Scripts\Package-Native.ps1:22 char:9
+         $wc.DownloadFile($source, $localFilename)
+         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    + CategoryInfo          : NotSpecified: (:) [], MethodInvocationException
    + FullyQualifiedErrorId : WebException

for
DownloadFile https://registrationcenter-download.intel.com/akdlm/irc_nas/18444/l_onemkl_p_2022.0.1.117_offline.sh "$PSScriptRoot/../packages/l_onemkl_p_2022.0.1.117_offline.sh"

[atlemann]

I'm currently trying to create C# methods for calling the DftiCreateDescriptor in a library, since I need single precision 3D FFT and thought it was most flexible to just expose all the Dfti methods in C# and perhaps add them to this project. But I just can't get it to work in my lib. Every time my code hits the DftiCreateDescriptor from C#, it crashes immediately with a 132 exit code.

[TSerious]

@atlemann you can checkout my repo: oneMKL.FFT.NET
It has most of the Dfti methods exposed.

[atlemann]

Looks very similar to what I tried. Perhaps it's some nuance with NixOS. Looks like you're targeting Windows?

[TSerious]

Yes, that's right.
I already thought about integrating what I did into MKL.NET, but the last time I tried I was not able to get it compiling (see the first post in this request).

[atlemann]

Looks like some parts of the FFTW api's are available, e.g. fftwf_plan_dft_3d which gives you a pointer to a plan for single precision 3D FFT. https://www.intel.com/content/www/us/en/docs/onemkl/developer-reference-fortran/2023-0/using-fftw3-wrappers.html

So you can define these:

private const int FFTW_FORWARD = -1;
private const int FFTW_BACKWARD = 1;
private const uint FFTW_ESTIMATE = 64;

[DllImport(Mkl.DllFile, CallingConvention = CallingConvention.Cdecl, ExactSpelling = true)]
static extern IntPtr fftwf_plan_dft_3d(int nx, int ny, int nz, void* @in, void* @out, int sign, uint flags);

[DllImport(Mkl.DllFile, CallingConvention = CallingConvention.Cdecl, ExactSpelling = true)]
static extern void fftwf_execute(IntPtr plan);

[DllImport(Mkl.DllFile, CallingConvention = CallingConvention.Cdecl, ExactSpelling = true)]
static extern void fftwf_destroy_plan(IntPtr plan);

and call it like this:

private static void Compute3D(void* matrixData, int nx, int ny, int nz, bool forward)
{
    var plan = IntPtr.Zero;
    try
    {
        int sign = forward ? FFTW_FORWARD : FFTW_BACKWARD;
        plan = fftwf_plan_dft_3d(nx, ny, nz, matrixData, matrixData, sign, FFTW_ESTIMATE); // FFTW_ESTIMATE is ignored, according to the docs
        if (plan == IntPtr.Zero)
            throw new InvalidOperationException("Failed to create FFTW plan");

        fftwf_execute(plan);

        // Scale for inverse FFT
        if (!forward)
        {
            var scale = 1.0f / ((float)nx * ny * nz);
            // Unsafe, since I've already fixed matixData outside this function
            Blas.Unsafe.scal(nx * ny * nz * 2, scale, (float*)matrixData, 1);
        }
    }
    finally
    {
        if (plan != IntPtr.Zero)
            fftwf_destroy_plan(plan);
    }
}