Leveraging SIMD Vectorization
With the advent of column store databases, there was an urge to make use of SIMD vector processing. It naturally fits into the way table data is arranged. Let's first briefly check what is SIMD. It stands for Single Instruction Multiple Data. Today, CPU instructions support this kind of mechanism where the same instruction can be executed simultaneously on multiple data elements. E.g. Say, you want to double all the column values. Or remove the red component of the RGB values of pixels of an image. For large data, these operations are CPU bottlenecks. So SIMD cuts the CPU time significantly by operating simultaneously on 2, 4, 8, 16 or 32 (or more) data elements depending on the size of each data element. So suppose we want to do "arr[i] *= 2" for each element of "int32 arr[]". Normally we would iterate through each of the elements for doing this operation. In the generated assembly code, MUL instruction will be run on each of the elements. With SIMD, we would