LDC dev here. All the credit for this goes to Ilya Yaroshenko for the expertly crafted implementation and the LLVM developers for efficient low-level code generation (inlining, register allocation and so on, and in some places auto-vectorization). LDC only needs to make sure not to "mess up" things too much.

D does play a significant role in this achievement, though – D's very powerful yet easy to use features for generics and introspection make it possible to finely tune the code for different parameters (sizes/dimensions/…) while still being easy to understand and modify.