Why Not?

Each parallel architecture defines a new programming model
Each programming model has its own optimization constrains
Different constraints require different algorithms for efficiency
No automatic way to write new algorithms