Parallelized implementation of 2 dimensional Fast Fourier Transform.
Problem Statement Compute a Fourier Transform of a given square matrix using the following methods:
- Discrete Fourier transform using threads on CPU
- Cooley-Tukey algorithm using Message Passing Interface (MPI) on CPU
- Cooley-Tukey algorithm using CUDA on GPU
Solution The threading was done using the threading library of C++. The matrix number of rows in the matrix was divided by 8 threads, each thread then ran the equations given for all of its allotted rows. The results were all stored in a new array, and the code waited for all threads to finish. The new array was then transposed, and the threads repeated the process on what was now the columns of the original matrix. The result was stored in the original array, which was then transposed once again to give the final result.
The MPI implementation of Danielson-Lanczos method uses recursion in order to compute the FastFourier Transform (FFT). The FFT was calculated by computing the DFT of the even positions, then the DFT of the odd positions in a row and then summing them. This method is implemented by first computing the FFT for the rows assigned to each rank, sending the completed rows to a single rank, which then transposes the matrix and repeats the process.
The CUDA implementation leverages parallel high performance computing to calculate the 2D DFT of an input matrix by computing the 1D DFT’s simultaneously. We make use of the parallelism offered by the blocks and the synchronism offered by the threads to achieve an optimal implementation.