DESIGN OF THE BACK-END PROCESSING SYSTEM FOR THE ACCELERATION OF THE MICROWAVE IMAGING RECONSTRUCTION ALGORITHMS

Microwave imaging (MWI) is a diagnostic tool whose working principle relies on the dielectric contrast between lesions and healthy tissues, and could be used, for instance, to detect breast cancer or brain strokes. This work aims to improve the speed of the processing of MWI data acquired when using a finite element contrast source inversion method. The processing consists of solving a large sparse and complex linear system with 24 right-hand sides. Multiple open-source solvers (UMFPACK, KLU, Eigen and MUMPS) were tested to solve this linear system, including direct and indirect methods, and the precision obtained from each trial was compared. The only direct solver tested that could solve the all the MWI linear systems of interest was MUMPS. The indirect methods tested did not achieve precise results. The original proposal of this thesis was to perform the factorization of the matrix and then use a GPU to accelerate the solution of the triangular linear systems, MUMPS does not support exporting the factorization results, though. Therefore, UMFPACK was used to factorize matrices from the SuiteSparse Collection in order to test OpenCL implementations in a GPU so to employ parallelism. Two different types of kernels to solve triangular linear systems were implemented: a) column block algorithms, which only worked for very small matrices and presented synchronization issues for bigger matrices; b) solving multiple right-hand sides in parallel. When solving multiple right-hand sides in parallel, the row-compressed format could be executed faster than the column-compressed format. The latter format had worst performance because it required more accesses to the GPU’s global memory. Two modified versions of the kernel to deal with column-compressed format using local memory were also implemented, which resulted in smaller run times, but still slower than executing the same task in a CPU. In general, the observed overhead of moving data to and from the GPU was greater than the time to execute the same task sequentially in a CPU for the tested matrices.