Enable MKL parallelism for FFT/iFFT

vlsergey — Fri, 13 Mar 2009 23:19:58 GMT

Hi,

I'm tryingusing MKL 10.1 to perforf FFT / iFFT transform. But, so far, according to process explorer and timing, only one processor core is used. SettingMKL_DYNAMIC=false, MKL_NUM_THREADS=4 has no effect.

Intel Core i7 920 @ 2.67 Ghz, Memory (RAM): 6 Gb

Windows Vista 64-but Enterprise

Intel MKL 10.1, mkl_java_stubs compiled by MS Visual Studio 2008 (with /openmp, using em64t libraries and internal compiler)

Initialization of DFTI descriptor is:

		desc_handle = DFTI.CreateDescriptor(DFTI.DOUBLE, DFTI.REAL,
				1, length);
		DFTI.SetValue(desc_handle, DFTI.PLACEMENT, DFTI.INPLACE);
		DFTI.CommitDescriptor(desc_handle);

Tried both in-place and out-place transformations. Data length are always power of two.

Thanks for any ideas, how to enable internal MKL parallelism.

Re: Enable MKL parallelism for FFT/iFFT

Dmitry_B_Intel — Tue, 17 Mar 2009 12:36:53 GMT

Hello, vlsergey

One-dimensional real-to-complex FFTs are not threaded currently. However, should you compute several transforms in a row, MKL will do it in parallel if you set DFTI_NUMBER_OF_TRANSFORMS > 1.

Dima

topic Enable MKL parallelism for FFT/iFFT in Intel® oneAPI Math Kernel Library

Enable MKL parallelism for FFT/iFFT

Re: Enable MKL parallelism for FFT/iFFT