I performed performance tests for the stencil operator for both single time instance and multiple iteration cases. I looked at the effect both grid size and number of iterations had on the overall performance (number of cycles per DOF). However, because I couldn't figure out how the vectorization was implemented, I was unable to propose and analytical performance model for the stencil operator. However, I did observe that with both increasing grid size, and increasing number of iterations, overall performance increased, suggesting that within the sizes of problems tested here, vectorization helped overcome the memory bandwidth problem.
I performed performance tests for the stencil operator for both single time instance and multiple iteration cases. I looked at the effect both grid size and number of iterations had on the overall performance (number of cycles per DOF). However, because I couldn't figure out how the vectorization was implemented, I was unable to propose and analytical performance model for the stencil operator. However, I did observe that with both increasing grid size, and increasing number of iterations, overall performance increased, suggesting that within the sizes of problems tested here, vectorization helped overcome the memory bandwidth problem.