FFTW.fft() fails to calculate spectrum

[StefanPofahl]

Hi,

yesterday I observed in one measurement that fft-fails to calculate the spectrum of my data.
The data is a nice sinus. I do not know what the root cause is.
The result is the same on Julia-LTS and Julia v1.8.0.
If I use SciPy.fft.fft() instead, the spectrum is calculated as expected.
I would like to send you the data, to enable debugging on your side.
Unfortunatly the file size is above the accepted file size,
can you offer a way to send the data?

Regards,

Stefan

FFTW_Debug_figures.zip

[StefanPofahl]

I was able to reproduce the error with syntetic data, here is my script:

using PlotlyJS, Printf
using SciPy, FFTW

sampling_rate = 250000
window_length = 47499754
frequ1        = 0.1
ampl1         = 2.1
phase1        = 0.0

# --- time and data vector:
vec_t    = collect(range(0, step = 1.0 / sampling_rate, length = window_length)) 
td_data_ = ampl1 .* sin.( (2 * pi * frequ1 .* vec_t) .+ deg2rad(phase1))

# --- relevant range in fft-result (meaningful bins):
bins_range = range(1, step = 1, stop = ceil(Int, window_length / 2))

# --- Frequncy Vector --------------------------------------------------------------------------------------
t_win_duration = window_length / sampling_rate
frequ_vec = 1 / t_win_duration .* (0 : (ceil(Int, window_length / 2) - 1))
frequ_vec = frequ_vec[2:end]

# --- FFTW -------------------------------------------------------------------------------------------------
amplitudes_FFTW = abs.(FFTW.fft(td_data_))[bins_range]
amplitudes_FFTW = amplitudes_FFTW .* (2/window_length)
amplitudes_FFTW = amplitudes_FFTW[2:end]

# --- SciPy -------------------------------------------------------------------------------------------------
amplitudes_SciPy = abs.(SciPy.fft.fft(td_data_))[bins_range]
amplitudes_SciPy = amplitudes_SciPy .* (2/window_length)
amplitudes_SciPy = amplitudes_SciPy[2:end]

# --- Plot --------------------------------------------------------------------------------------------------
plt_range = range(1, length = 40)
hdl1 = PlotlyJS.scatter(; x = frequ_vec[plt_range], y = amplitudes_FFTW[plt_range],  name = "FFTW")
hdl2 = PlotlyJS.scatter(; x = frequ_vec[plt_range], y = amplitudes_SciPy[plt_range], name = "SciPy")
hdl_plt = PlotlyJS.Plot([hdl1, hdl2])

[StefanPofahl]

The problem occurs if I turn on the MLK by this command: FFTW.set_provider!("mkl")
If I switch back to fftw by the command: FFTW.set_provider!("fftw") the amplitude signal can be calculated correctly.
If the input is real and I can use: rfft() instead of fft() the problem can also so be prevented.

[stevengj]

If it's a bug in MKL, there's nothing I can do about it — go complain to Intel. (They will probably want you to reproduce it in a minimal C program, however.)

[StefanPofahl]

[StefanPofahl]

They have an open github repository, including the functionality to open issues, I have opened a ticket:
oneapi-src/oneMKL#222

[StefanPofahl]

I was informed that this group was not correct, therefore, I opened a new issue in another group:
[oneAPI Math Kernel Library](https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/function-fft-sometimes-fails-to-calculate-result/m-p/1414125#M33641)

Let's see, if this is the right place to inform Intel about this bug.