Hello,
I need to compute the 1d convolution. Intel MKL offers two basic strategies to do this.
1) Explicit implementation of the convolution theorem by the user i.e. perform DFT's on the input data and on the kernel. Multiply the results in the Fourier domain element wise. Next perform an inverse DFT to get the desired result.
2) Compute the convolution directly by using VSL math function in FFT mode.
To my knowlegde the computation using a VSL function is not internally parallelized by OpenMP. Hence it does not benefit from multicore architectures right ? But the FFT functions within MKL are indeed internally parallelized by OpenMP and I can even call the cluster version for very huge problems. My question is know what's the most effective way to calculate the convolution the first or the second way ? The Fourier transform of the kernel might be known analytically (since the kernel is given by a known function). Is there a way to exploit this in the second approach ? Within the first approach calculation of the FFT of the kernel is avoided.
Best regards,
Felix