The following sample program shows how to employ internal OpenMP* threading in Intel® oneAPI Math Kernel Library (oneMKL) for FFT computation.
To specify the number of threads inside Intel® oneAPI Math Kernel Library (oneMKL), use the following settings:
set MKL_NUM_THREADS = 1 for one-threaded mode;
set MKL_NUM_THREADS = 4 for multi-threaded mode.
/* C99 example */
#include "mkl_dfti.h"
float data[200][100];
DFTI_DESCRIPTOR_HANDLE fft = NULL;
MKL_LONG dim_sizes[2] = {200, 100};
/* ...put values into data[i][j] 0<=i<=199, 0<=j<=99 */
DftiCreateDescriptor(&fft, DFTI_SINGLE, DFTI_REAL, 2, dim_sizes);
DftiCommitDescriptor(fft);
DftiComputeForward(fft, data);
DftiFreeDescriptor(&fft);
The following Example “Using Parallel Mode with Multiple Descriptors Initialized in a Parallel Region” and Example “Using Parallel Mode with Multiple Descriptors Initialized in One Thread” illustrate a parallel customer program with each descriptor instance used only in a single thread.
Specify the number of OpenMP threads for Example “Using Parallel Mode with Multiple Descriptors Initialized in a Parallel Region” like this:
set MKL_NUM_THREADS = 1 for Intel® oneAPI Math Kernel Library (oneMKL) to work in the single-threaded mode (recommended);
set OMP_NUM_THREADS = 4 for the customer program to work in the multi-threaded mode.
Note that in this example, the program can be transformed to become single-threaded at the customer level but using parallel mode within Intel® oneAPI Math Kernel Library (oneMKL). To achieve this, you must set the parameter DFTI_NUMBER_OF_TRANSFORMS = 4 and to set the corresponding parameter DFTI_INPUT_DISTANCE = 5000.
/* C99 example */
#include "mkl_dfti.h"
#include <omp.h>
#define ARRAY_LEN(a) sizeof(a)/sizeof(a[0])
// 4 OMP threads, each does 2D FFT 50x100 points
MKL_Complex8 data[4][50][100];
int nth = ARRAY_LEN(data);
MKL_LONG dim_sizes[2] = {
ARRAY_LEN(data[0]),
ARRAY_LEN(data[0][0])
}; /* {50, 100} */
int th;
/* ...put values into data[i][j][k] 0<=i<=3, 0<=j<=49, 0<=k<=99 */
// assume data is initialized and do 2D FFTs
#pragma omp parallel for shared(dim_sizes, data)
for (th = 0; th < nth; ++th)
{
DFTI_DESCRIPTOR_HANDLE myFFT = NULL;
DftiCreateDescriptor(&myFFT, DFTI_SINGLE, DFTI_COMPLEX, 2, dim_sizes);
DftiCommitDescriptor(myFFT);
DftiComputeForward(myFFT, data[th]);
DftiFreeDescriptor(&myFFT);
}
Specify the number of OpenMP threads for Example “Using Parallel Mode with Multiple Descriptors Initialized in One Thread” like this:
set MKL_NUM_THREADS = 1 for Intel® oneAPI Math Kernel Library (oneMKL) to work in the single-threaded mode (obligatory);
set OMP_NUM_THREADS = 4 for the customer program to work in the multi-threaded mode.
/* C99 example */
#include "mkl_dfti.h"
#include <omp.h>#
define ARRAY_LEN(a) sizeof(a)/sizeof(a[0])
// 4 OMP threads, each does 2D FFT 50x100 points
MKL_Complex8 data[4][50][100];
int nth = ARRAY_LEN(data);
MKL_LONG dim_sizes[2] = {
ARRAY_LEN(data[0]),
ARRAY_LEN(data[0][0])
}; /* {50, 100} */
DFTI_DESCRIPTOR_HANDLE FFT[ARRAY_LEN(data)];
int th;
/* ...put values into data[i][j][k] 0<=i<=3, 0<=j<=49, 0<=k<=99 */
for (th = 0; th < nth; ++th)
DftiCreateDescriptor(&FFT[th], DFTI_SINGLE, DFTI_COMPLEX, 2, dim_sizes);
for (th = 0; th < nth; ++th)
DftiCommitDescriptor(FFT[th]);
// assume data is initialized and do 2D FFTs
#pragma omp parallel for shared(FFT, data)
for (th = 0; th < nth; ++th)
DftiComputeForward(FFT[th], data[th]);
for (th = 0; th < nth; ++th)
DftiFreeDescriptor(&FFT[th]);
The following Example “Using Parallel Mode with a Common Descriptor” illustrates a parallel customer program with a common descriptor used in several threads.
#include "mkl_dfti.h"
#include <omp.h>
#define ARRAY_LEN(a) sizeof(a)/sizeof(a[0])
// 4 OMP threads, each does 2D FFT 50x100 points
MKL_Complex8 data[4][50][100];
int nth = ARRAY_LEN(data);
MKL_LONG len[2] = {ARRAY_LEN(data[0]), ARRAY_LEN(data[0][0])};
DFTI_DESCRIPTOR_HANDLE FFT;
int th;
/* ...put values into data[i][j][k] 0<=i<=3, 0<=j<=49, 0<=k<=99 */
DftiCreateDescriptor(&FFT, DFTI_SINGLE, DFTI_COMPLEX, 2, len);
DftiCommitDescriptor(FFT);
// assume data is initialized and do 2D FFTs
#pragma omp parallel for shared(FFT, data)
for (th = 0; th < nth; ++th)
DftiComputeForward(FFT, data[th]);
DftiFreeDescriptor(&FFT);