What is the proper way to zero-pad a 2d data and perform backward transformation?

raavi — Thu, 06 Jun 2024 11:38:09 GMT

I am trying to understand the correct way to zero-pad the 2d-data.

This way I did not get the correct output.

MKL_LONG dim_child[2]; dim_child[0] = 9; dim_child[1] = 17; MKL_LONG dim_parent[2]; dim_parent[0] = 12; dim_parent[1] = 23; MKL_LONG status; DFTI_DESCRIPTOR_HANDLE frwd_hndl = NULL; DFTI_DESCRIPTOR_HANDLE bkwd_hndl = NULL; double forward_scale = sqrt(2.0 * pi) / static_cast<double>(dim_child[0] * dim_child[1]); status = DftiCreateDescriptor(&frwd_hndl, DFTI_DOUBLE, DFTI_COMPLEX, 2, dim_child); status = DftiSetValue(frwd_hndl, DFTI_FORWARD_SCALE, forward_scale); status = DftiCommitDescriptor(frwd_hndl); status = DftiCreateDescriptor(&bkwd_hndl, DFTI_DOUBLE, DFTI_COMPLEX, 2, dim_parent); status = DftiCommitDescriptor(bkwd_hndl); int size_child = dim_child[0] * dim_child[1]; int size_parent = dim_parent[0] * dim_parent[1]; MKL_Complex16 *fourier_coeff = NULL; MKL_Complex16 *filterd_coeff = NULL; fourier_coeff = (MKL_Complex16*)mkl_malloc(size_child * sizeof(MKL_Complex16), 64); filterd_coeff = (MKL_Complex16*)mkl_malloc(size_parent * sizeof(MKL_Complex16), 64); for (int j=0; j<lg_child; j++) { for (int i=0; i<lt_child; i++) { int index = i + (j*lt_child); fourier_coeff[index].real = // some_value; fourier_coeff[index].imag = // some_value; } } status = DftiComputeForward(frwd_hndl, fourier_coeff); status = DftiFreeDescriptor(&frwd_hndl); for (int j_prime=0; j_prime<lg_parent; j_prime++) { for (int j=0; j<lg_child; j++) { for (int i=0; i<lt_child; i++) { int index = i + (j * lt_child); int index_filter = i + (j_prime * lt_parent); // zero-padding at the end of each row of length (lt_parent - lt_child) // [x x x x x x x x x x x x x x x x x 0 0 0 0 0 0 x x x x x x x x x x x x x x x x x 0 0 0 0 0 0...] // _______________17________________ _____6_____ filterd_coeff[index_filter].real += (fourier_coeff[index].real); filterd_coeff[index_filter].imag += (fourier_coeff[index].imag); } } } status = DftiComputeBackward(bkwd_hndl, filterd_coeff); status = DftiFreeDescriptor(&bkwd_hndl); for (int j=0; j<lg_parent; j++) { for (int i=0; i<lt_parent; i++) { int index = i + (j*lt_parent); std::cout << filterd_coeff[index].real<<"\t"<<filterd_coeff[index].imag<<std::endl; } } mkl_free(fourier_coeff); mkl_free(filterd_coeff);

Re:What is the proper way to zero-pad a 2d data and perform backward transformation?

Fengrui — Mon, 24 Jun 2024 20:44:12 GMT

Hi,

Thank you for posting in the forum!

I'm trying to understand the issue here. It looks the code is trying to zero-pad the Fourier coefficients after the FFT. Could you please provide more information about what this zero-padding is trying to achieve and what output is expected?

Thanks,

Fengrui

topic Re:What is the proper way to zero-pad a 2d data and perform backward transformation? in Intel® oneAPI Math Kernel Library

What is the proper way to zero-pad a 2d data and perform backward transformation?

Re:What is the proper way to zero-pad a 2d data and perform backward transformation?