VecTcl

Memory bandwidth benchmark

Adding to large arrays is mostly limited by memory bandwidth; this microbenchmark compares the speed of various operations against memcpy and a simple C-coded loop that adds to vectors. memcpy outperforms the C-coded loop by a factor of 2. VecTcl assignment operators (i.e., x+=y) should ideally approach the C-coded loop. For vectors, we are almost there. Tall matrices are worst. Overall, reductions (i.e., sum(x)) score lowest.