I would first learn a couple of Array Programming Languages, first a mainstream one, e.g. Julia/R/Matlab/Octave, and then one from the APL-family (APL/J/K/Q/etc.)
Lear how to write algorithms using vectorization only (without if-s and for-s).
Look into OpenCL / CUDA programming models.
Then it will be much easier to learn native SIMD programming for specific ISAs.